Ahmet Halit
Ünsal
Software Manager & Full-Cycle Product Architect.
15+ years spanning safety-critical embedded systems, large-scale team management, and zero-to-production product engineering.
Independent Ventures
Products conceived, architected, built, and operated entirely as solo ventures. Every product, business, infrastructure, and operational decision is owned by me.
Project AURA
Automated Unbiased Return Algo — a production-grade, 24/7 algorithmic trading system for liquid crypto markets. Runs against the live Binance Spot exchange from a self-hosted ARM64 edge node. Designed around deterministic execution, persistent state, and post-incident hardening rather than throughput.
Technical Architecture
- Engineered a fully synchronous execution engine to guarantee single-threaded determinism across the data → asset-selection → risk → portfolio → broker pipeline, eliminating an entire class of race conditions inherent in async/event-loop trading stacks.
- Selected synchronous over async by design — the system is I/O-bounded by exchange rate limits, not CPU-bounded;
time.sleep()-paced retry loops produce a strictly auditable execution trace and a single source of truth for portfolio state at any instant. - Designed a hot-reloadable configuration plane (
Cfg/config_overrides.json) consumed via dataclass-backedCFGsnapshots, enabling tuning of risk thresholds, regime parameters, and universe filters between cycles with no service restart and no Python module reimport. - Built a dual-service topology — a live trading loop and a walk-forward parameter optimizer — communicating through a single JSON IPC surface, so optimization output flows into live execution with one atomic write.
- Implemented a 4-stage successive-halving walk-forward optimizer (ETA=3 reduction, budget-windowed evaluation) that tunes selector, regime, risk, and defensive-overlay knobs against rolling historical windows; scores candidates on a tuple of
(wins, avg_sharpe, avg_excess, avg_aura)to favor robust generalization over single-window overfit. - Architected SQLite-backed persistence across
events,intents,orders,snapshots,heartbeats, and akvstate store, providing crash-safe reconciliation: any broker-acknowledged order whose fill callback is lost is refreshed against the exchange on the next service boot. - Hardened the broker reconciliation path with a sync-failure guard: when a transient
recvWindowor network error returns an empty exchange position set, the system flags valuation as suspect and skips PM reconciliation that tick rather than wiping in-memory state — eliminating a class of phantom kill-switch trips driven by API blips. - Implemented entry-price recovery from persisted snapshots so that, after a restart, stop-loss enforcement is restored against the original cost basis instead of resetting to current market price.
- Coded a 2-cycle hysteresis layer over raw BULL/SIDEWAYS/BEAR regime detection (
BTCtotal-return vs configurable threshold over an N-day lookback), requiring N consecutive confirmations before a regime transition commits — suppressing daily regime jitter at the boundary. - Built a custom multi-provider market-data layer with a coverage-aware provider chain (Binance Vision → Tiingo → Polygon → Alpha Vantage → yfinance → CoinGecko), per-symbol Parquet caching, read-through fallback dirs, freshness short-circuits, and provider lookback padding — cutting "possibly delisted" cascade failures and accelerating walk-forward backtests by orders of magnitude.
Infrastructure & Deployment
- Deployed the production stack on a Raspberry Pi 5 (Ubuntu/Debian ARM64) as two long-running
systemdservices —trading.serviceandpropagate.service— running continuously withjournalctl-based observability and Telegram notification on every order, regime change, and kill-switch event. - Engineered for edge-computing reliability — low-power footprint (≈ 15 W steady-state), local SSD persistence, and full operational autonomy without dependence on cloud compute; the node has run multi-week stretches across regional power blips with no operator intervention.
- Constrained BLAS thread pools to 1 (
OMP_NUM_THREADS,OPENBLAS_NUM_THREADS,MKL_NUM_THREADS) at process boot to prevent NumPy/pandas oversubscription on ARM cores during 600+ symbol walk-forward sweeps. - Designed startup invariants — every service boot performs (a) persistence reconciliation against the exchange, (b) PM-to-broker position sync, (c) entry-price recovery from snapshots, and (d) kill-switch state rehydration from the
kvstore before the first trade can fire. - Instrumented permanent diagnostic logging (
[RISK_TICK],[SL_CHECK],[BOOT],[SYNC_GUARD],[ENTRY_RECOVER]) — adds ~3 lines/min in steady state and makes silent regressions impossible to miss fromjournalctlalone. - Engineered a CCXT-mediated Binance Spot integration with sell-clamp retry logic, BNB fee reserve management, persistent
clientOrderIdhashing for idempotent replays, and reconciliation passes that mark unreachable ordersRECONCILE_FAILEDrather than leaving them stranded as live pending work. - Hardened the kill-switch liquidation path to retry partial liquidations on network instability — converting a prior single-attempt sell sequence into a verified loop that re-checks exchange positions and re-issues sells until the portfolio is flat.
Strategic Portfolio Management
Operating role: Portfolio Manager. Full P&L responsibility on live capital. Authority over risk parameters, universe construction, capital-utilization caps, regime thresholds, and the optimizer search space. Every code merge runs against real funds at the next hot-reload cycle.
- Defined and tune the "Unbiased Return" thesis — a control-oriented framework that enforces strict adherence to pre-committed parameters (position stability gates, cost-benefit thresholds, regime-aware utilization caps) so that no single trade is influenced by discretionary sentiment or recency bias.
- Manage a multi-asset crypto universe anchored on a high-liquidity whitelist (BTC, ETH, BNB, XRP) and a dynamic momentum-screened tail sourced daily from Binance via multi-horizon (6m/12m/24m) return scoring, drawdown caps, and SMA-200 trend filters.
- Curate the defensive overlay: in SIDEWAYS / BEAR regimes the system routes a configurable fraction of investable capital into Pax Gold (PAXG-USD) as a tokenized-gold safe haven, with a BTC-USD fallback; capital deployment is further governed by regime-specific utilization caps to deliberately retain dry powder during low-conviction tape.
- Author the composite scoring function that ranks the universe — a weighted blend of
trend_blend_alpha-mixed long/short-horizon momentum, Sharpe, max-drawdown depth, and an optional news component — with a regime-multiplier overlay that amplifies momentum in BULL and penalizes volatility in BEAR. - Operate a multi-layer risk stack: per-position stop-loss and take-profit (with persistent entry-price recovery), a rolling-loss kill-switch with configurable baseline window and cooldown, a portfolio-level cost-benefit gate that suppresses uneconomical rebalances, and an overlap-plus-score-improvement gate that suppresses turnover when conviction has not materially changed.
- Iterate the optimizer search space — define the propagation grid (regime threshold, trend-blend mixing, stop-loss bands, take-profit triggers, kill-switch sensitivity, defensive weights, utilization caps) and review walk-forward output before promoting any knob change to live capital.
- Performed root-cause analysis on every live incident (phantom kill-switch trips, stop-loss reachability regressions, lost fill callbacks, broker timestamp drift, universe screen blackouts during bear regimes) and merged structural fixes — not workarounds — back into the system with verification gates that surface any future recurrence in
journalctlimmediately. - Continuously refine the Unbiased Return logic to balance defensive preservation against momentum capture across regime transitions, tightening the gap between propagation backtest yield and live realized return.
Stack: Python 3 (synchronous), pandas, NumPy, CCXT, SQLite, systemd, journalctl, custom Walk-Forward engine, Binance Spot REST/WebSocket via CCXT, Telegram bot API.
Deployment surface: ARM64 Raspberry Pi 5, Ubuntu/Debian; self-hosted, no cloud dependency.
Operational mode: Live capital, 24/7, single-operator portfolio management.
Professional Experience
15+ years of driving engineering excellence.
AHMET HALIT UNSAL
Senior Engineering Manager & Full-Cycle Product Developer
📍 Izmir, TR | ✉️ aunsal89@gmail.com | 📞 +90-552-232-8144 🔗 linkedin.com/in/ahmethalitunsal | 🐙 github.com/aunsal89
Profile
Engineering leader with 15+ years spanning safety-critical embedded systems, large-scale team management, and full-cycle product development. Directed a 30-engineer organization delivering ISO 26262-compliant software for EV powertrains at production scale, while concurrently architecting and operating two independent ventures from zero to production — a clinical-grade mobile application and a 24/7 live-capital algorithmic trading system. Equally fluent in organizational design, control algorithm development, React Native product engineering, and quantitative portfolio management. Known for owning outcomes end-to-end: root-causing production failures, shipping structural fixes, and building systems that harden against recurrence.
Technical Toolbox
| Domain | Technologies & Standards |
|---|---|
| Embedded & Automotive | C/C++, FreeRTOS, AUTOSAR BSW/ASW, ISO 26262 (ASIL), SPICE, CAN/LIN/Ethernet, MATLAB/Simulink, Stateflow, Lauterbach Trace32, dSpace MABX |
| Control Systems | PMSM/IPMSM motor control, MTPA, flux weakening, PID synthesis, traction control, powertrain state machines, HIL/MiL simulation |
| Product Engineering | React Native (Expo SDK 54), TypeScript, TanStack Query v5, Supabase (Postgres + RLS + Edge Functions), Vite, React 19, NativeWind, Expo Router, EAS Build/Submit |
| Backend & Infrastructure | PostgreSQL, Row-Level Security, Deno Edge Functions, Resend, SQLite, systemd, journalctl, CCXT, Binance Spot REST/WebSocket |
| Quantitative / Algo | Python 3 (synchronous), pandas, NumPy, walk-forward optimization, regime detection, multi-layer risk management, Sharpe/drawdown-based scoring |
| Engineering Operations | Git, SVN, JIRA, CI/CD pipelines, App Store Connect, EAS Build, TestFlight, SPF/DKIM/DMARC, Telegram bot API |
| Methodologies | SAFe / Scrum, Situational Leadership (SL-II), Model-Based Development, V-Model, data-driven decision-making |
Education
MSc, Advanced Control Theory | Istanbul Technical University | 2014 – 2016
BSc, Electrical & Computer Engineering | Iowa State University | 2007 – 2011
Professional Experience
Software Manager | NISO TECHNOLOGY CO
May 2021 – Present | Izmir, TR
Directed a multidisciplinary engineering organization of 30 software engineers across 4 team leads, delivering mission-critical embedded software for EV powertrain systems including ECUs, traction inverters, and on-board chargers (OBCs).
- Scaled organizational capability by establishing a team-lead layer, defining ownership boundaries, and running structured performance reviews and mentorship programs — reducing manager single points of failure and enabling parallel delivery across subsystems.
- Enforced ISO 26262 and SPICE compliance across BSW and ASW layers developed on FreeRTOS and other RTOS platforms, ensuring functional safety traceability from requirements through integration.
- Deployed CI/CD pipelines and automated tooling that compressed release cycle times and shifted regression detection left, reducing late-stage integration failures.
- Led UN R155/R156 cybersecurity standard implementation activities, aligning software development processes with vehicle cybersecurity management system (CSMS) requirements.
- Designed and deployed C-based middleware solutions for high-reliability embedded applications, and contributed directly to diagnostic and safety-critical algorithms for powertrain systems.
- Instituted Git-based version control strategies with branching policies that enforced traceability and streamlined cross-team collaboration across a large, concurrent codebase.
- Drove data-driven project management by deploying analytics tooling to surface velocity, defect density, and milestone risk metrics to stakeholders.
- Managed resource allocation and risk mitigation across competing program milestones, acting as the primary interface between software, hardware, and systems engineering teams.
Team Leader | ECEMTAG
November 2017 – May 2021 | Istanbul, TR
Led design and validation of model-based control algorithms for ECUs and TCUs across automotive and defense programs.
- Introduced automated testing frameworks and HIL simulation environments that raised software verification coverage and reduced physical prototype dependency.
- Developed safety and redundancy functions in compliance with military standards, coordinating requirements directly with defense customers.
- Acted as cross-functional integration lead, aligning software and hardware milestones across embedded, systems, and validation teams using JIRA-based program tracking.
- Contributed to E/E architecture process development and rollout, including tool qualification and development workflow standardization.
Lead Researcher | TUBITAK (Scientific and Technological Research Council of Turkey)
December 2016 – November 2017 | Kocaeli, TR
- Directed automotive and defense research programs focused on advanced control systems and diagnostics.
- Developed a product-level tool for data dictionary generation, calibration, and MiL (Model-in-Loop) testing, accelerating algorithm validation workflows.
- Characterized IPMSM performance in motoring and regenerative modes on a dynamometer test bench.
Senior Systems Engineer | TUMOSAN
January 2016 – December 2016 | Istanbul, TR
- Developed automatic shift-selection logic using Simulink Stateflow for a hybrid drivetrain platform.
- Designed traction control algorithms for slip-slide protection under variable surface conditions.
- Optimized fuel consumption and battery usage profiles using multi-objective cost functions.
Systems Engineer | METRO CO
February 2013 – January 2016 | Kocaeli, TR
- Built a longitudinal vehicle model and derived drivetrain power and torque requirements from first principles.
- Developed a full PMSM driver including Maximum Torque Per Ampere (MTPA) and flux-weakening control algorithms.
- Automated PID tuning via phase and gain margin calculation, and conducted HIL validation using dSpace MABX-I.
Researcher | Iowa State University — Bio-Electric Laboratory
February 2011 – January 2013 | Ames, IA
- Supervised advanced research on transcranial magnetic stimulation (TMS) and electric field modeling.
- Instructed electric machines coursework to sophomore engineering students.
Independent Ventures
The following projects were conceived, architected, built, and operated entirely as solo ventures. Every product, business, infrastructure, and operational decision is owned by me.
EduTrace — Founder, Architect & Sole Engineer
2024 – Present
EduTrace (edutrace.net) is a clinical-grade special-education evaluation platform for licensed Turkish evaluators, covering diagnoses including ASD, ADHD, SLD, ID/DD, SI, OHI, and ED. The product spans a React Native mobile app, a Vite/React marketing site, a Supabase backend, and an admin evaluator-onboarding workflow. I own 100% of product strategy, UX, architecture, implementation, DevOps, App Store operations, infrastructure, compliance, and customer-facing communication.
Business & Product Ownership
- Defined the product vision, target market (KVKK-aware Turkish clinical evaluators), go-to-market strategy, pricing model, and App Store presence from inception.
- Registered and operate
edutrace.net, including DNS management, SPF/DKIM/DMARC configuration, and Resend-backed authenticated outbound email. - Managed the full App Store Connect workflow: app record, screenshots, metadata, TestFlight provisioning, and 6 production builds submitted across iterative releases.
- Operate all vendor relationships: Supabase project administration, Resend email routing, OpenRouter and Google Gemini AI APIs, and EAS Build CI configuration.
Technical Architecture Highlights
- Offline-first clinical workflow: evaluation sessions auto-save every 30 seconds locally, queue mutations under typed operations (
save_draft / finalize / update_completed), deduplicate bytype + sessionId, and flush atomically on network restoration — enabling reliable operation in clinic environments with intermittent Wi-Fi. - JSON-driven dynamic template system (M9): migrated evaluation templates from hard-coded TypeScript to a
templates.definition JSONB + version TEXTPostgres schema with a 3-tier fallback resolver (remote → bundled JSON → hard-coded default), enabling non-developer template updates with no app release required. - Fully bilingual clinical content pipeline: designed a
LocalizedString = string | { en: string; tr: string }discriminated type with aresolveLabel()helper threaded through every render and PDF export site, producing correct English and Turkish clinical reports from a single JSON definition. - AI study-plan engine with dual-provider resilience: implemented OpenRouter REST + Direct Gemini SDK with per-attempt exponential backoff with jitter (1 s → 2 s → 4 s + ≤250 ms), a model fallback chain of up to 10 models, and a transient-vs-hard-failure classifier — delivering 12-week S.M.A.R.T. study plans even under upstream rate-limiting.
- Admin onboarding with realtime state machine: architected a
profiles.statusmachine (pending → approved | rejected) with private Supabase Storage signed URLs (5-min TTL), sequential document capture (government ID + diploma + selfie), and realtime channels with per-instanceuseId()topic suffixes — enabling KVKK-compliant evaluator gating with live queue updates and zero polling. - Multi-tenant cache isolation (M15): identified and eliminated a critical cross-tenant data-leak where a prior user's student data briefly surfaced on account switch; implemented identity-change detection in
AuthContextthat firesqueryClient.clear()+offlineStorage.clearUserScopedCaches()on every identity transition. - Root-caused and fixed a high-stakes EAS build invariant crash (2026-05-09): diagnosed a React
19.1.0 → 19.2.6version bump that silencedERESOLVEduringnpm cibut crashed on device due to RN 0.81.5's renderer being compiled against a specific React version.
Stack: React Native (Expo SDK 54), TypeScript, Expo Router, TanStack Query v5, NativeWind v4, Supabase (Postgres + RLS + Edge Functions + Realtime + Storage), Vite + React 19, Deno, Resend, OpenRouter, Google Gemini, EAS Build/Submit, App Store Connect.
Project AURA (Automated Unbiased Return Algo) — Architect, Engineer & Portfolio Manager
Ongoing
AURA is a production-grade, 24/7 algorithmic trading system operating against the live Binance Spot exchange, deployed on a self-hosted ARM64 edge node. I am simultaneously the system architect, the SRE, and the portfolio manager with full P&L responsibility on live capital.
Architectural & Engineering Decisions
- Designed a fully synchronous execution engine — a deliberate architectural choice over async event-loop approaches. The system is I/O-bounded by exchange rate limits, not CPU-bounded; synchronous
time.sleep()-paced retry loops produce a strictly auditable, single-threaded execution trace and a single source of truth for portfolio state at any instant, eliminating an entire class of race conditions. - Deployed on Raspberry Pi 5 (ARM64, Ubuntu/Debian) as two
systemdlong-running services (trading.service,propagate.service) withjournalctl-based observability and Telegram notifications on every order, regime change, and kill-switch event — achieving full operational autonomy with ≈15 W steady-state power draw. - Implemented a 4-stage successive-halving walk-forward optimizer with ETA=3 reduction and budget-windowed evaluation across rolling historical windows; scores candidates on a
(wins, avg_sharpe, avg_excess, avg_aura)tuple to favor generalization over single-window overfit, communicating results to the live loop via a single JSON IPC surface. - Architected SQLite-backed persistence across
events,intents,orders,snapshots,heartbeats, and akvstate store, enabling crash-safe reconciliation: any broker-acknowledged order whose fill callback is lost is recovered against the exchange at the next service boot. - Engineered a sync-failure guard on the broker reconciliation path: when a transient API error returns an empty position set, the system flags valuation as suspect and skips portfolio manager reconciliation that tick — eliminating phantom kill-switch trips driven by exchange API blips.
- Built a multi-provider market-data layer (Binance Vision → Tiingo → Polygon → Alpha Vantage → yfinance → CoinGecko) with per-symbol Parquet caching and coverage-aware fallback, cutting cascade failures and accelerating walk-forward backtests by orders of magnitude.
Portfolio Management
- Operate a multi-asset crypto universe: a high-liquidity anchor whitelist (BTC, ETH, BNB, XRP) combined with a dynamic momentum-screened tail ranked by multi-horizon (6m/12m/24m) return scoring, drawdown caps, and SMA-200 trend filters.
- Author the composite scoring function: a weighted blend of trend-mixed momentum, Sharpe, max-drawdown depth, and a regime-multiplier overlay that amplifies momentum signals in BULL regimes and penalizes volatility in BEAR regimes.
- Operate a multi-layer risk stack: per-position stop-loss/take-profit with persistent entry-price recovery across restarts, a rolling-loss kill-switch with configurable baseline window and cooldown, a portfolio-level cost-benefit gate, and a position-overlap gate that suppresses uneconomical turnover.
- Performed root-cause analysis on every live incident — phantom kill-switch trips, stop-loss reachability regressions, lost fill callbacks, broker timestamp drift, universe screen blackouts — and merged structural fixes back into the system with permanent diagnostic logging that makes silent regressions impossible to miss.
Stack: Python 3 (synchronous), pandas, NumPy, CCXT, SQLite, systemd, journalctl, Binance Spot REST/WebSocket, Telegram Bot API, custom walk-forward engine. Deployment: ARM64 Raspberry Pi 5, self-hosted, no cloud dependency, live 24/7.