RaghavendraSriram
Sr. OBD Engineer · Expat @ DAF Trucks NV, Netherlands · MS EE + MBA · 9+ years.
automation tool @ Cummins
calibration owned end-to-end
CARB, EPA, Euro VI
OBD-critical software
Technical program lead.
Owns delivery end-to-end.
Sr. OBD calibration engineer at Paccar with 9+ years owning compliance-critical software - currently on a 2-year expat at DAF Trucks NV, Netherlands, rolling out misfire detection across PACCAR P14 platforms. Built automation that cut extraction time by 80% at Cummins. MY22-27 engine years owned end-to-end across CARB, EPA, Euro VI.
MS Electrical Engineering · MBA, UIUC 2026. Two peer-reviewed publications and one issued patent - listed below. Building AI governance platforms in parallel.
Programs I've owned
End-to-end delivery across regulatory, engineering, and cross-functional dimensions.
Career trajectory
What I build
13 shipped repositories spanning AI platforms, safety systems, financial intelligence, and systems research. Click any card for the full case study.
- Decomposes user intent into deterministic task graphs
- Routes to SLMs (temp=0) or Claude via YAML-driven pre/post-execution policies
- HMAC-SHA256-signed immutable audit trails; cost tracked per team
- No autonomous agents - deterministic execution only
- Local reverse proxy (Go 1.26) intercepting Claude Code at :7777
- 13-step pipeline: SHA256 dedup, bash-result caching (2h TTL), image stripping + OCR
- 4-tier memory manager + BadgerDB knowledge graph with hybrid ranking
- Measured 67% token reduction on 635-request session (295K -> 98K tokens)
- Local fine-tuning workbench for 135M-7B models (Python)
- Hardware-tier classification probes actual RAM/VRAM - refuses jobs that won't fit, no silent OOM
- MLX-LM primary on Apple Silicon, CUDA via accelerate launch, CPU fallback
- 30 RAM-tagged models, 6 export formats (GGUF/MLX/CoreML/ONNX/TFLite/Ollama), 101 unit tests
- Production Bloomberg alternative - 282 features across 30 tabs
- Real-time charting (35+ indicators in Web Worker), multi-LLM streaming (Claude 3.5/Gemini 2.0/GPT-4)
- 3,000-path Monte Carlo optimizer; institutional data (13-F/Form 4/Congress trades/dark pool)
- 5-guard self-learning trading bot + SHA-256 hash-linked audit chain; 335/335 tests, 0 TS errors
- YOLOv8m TensorRT FP16 + ByteTrack at 15 fps/camera (1,500 fps cluster on 3-4 RTX 4090s)
- Rolling 60-sec FFmpeg segment buffer for stream-copy clip extraction (<500ms)
- Ed25519 chain-of-custody signing for tamper-proof evidence
- OSHA Form 300/300A auto-export, DART/TRIR, ITA e-submission; 30+ API routers
- 25 open data sources (USGS, GDACS, NASA FIRMS, ACLED, aviation, maritime, weather, ocean) into unified PostGIS incidents
- Two-stage Claude pipeline - Haiku extraction then Sonnet narrative generation
- Atomic Redis token budgeting (490K/day, 45 summaries/day, 5/session) with layered degradation
- Geohash cache reuse; LibreTranslate for multilingual incident normalization
- 5 workspace crates (~5K LoC): no_std microkernel with SPSC ring bus
- Neural HAL (NPU/GPU runtime via JitPatchContext); USDF vector-addressed storage with VFS bridge
- Ring-0 agent swarm (Synthesis/Orchestrator/Curator/MemoryOptimizer/Janitor)
- SWIM-style gossip cluster with RDMA compute credits; not bootable - 28 issues documented
- 8 connectors: SAP OData v4 OAuth2, Ariba, CDK, Manhattan, Snowflake, TMS, Warranty, Demand
- 13 SQLAlchemy tables; APScheduler for 15-min ingest, 6am daily, Monday weekly jobs
- Z-score anomaly detection (60d lookback, z>=2 warning / z>=3 critical)
- 14-day demand forecast (40% trend + 60% MA); deployed on Render
- Decentralized P2P AI compute grid - developers pool local GPU hardware (Ollama/vLLM) into a unified async API gateway
- EIP-191 Ethereum wallet auth - guild admission verified on every WebSocket connect via admin-signed credentials
- Smart dispatch matches consumer requests to least-loaded node; 3-layer failure handling with APScheduler requeue sweep
- Compute credits settled hourly via PL/pgSQL batch - one transaction per hour instead of a write-lock per inference
- Intraday mean-reversion strategy driven entirely by Claude Code via a brokerage MCP server - no compiled trading engine
- Strategy config, session prompts, and cron runner; every decision logged to trade history and a strategy feedback log
- Phase-gated: signal validation (no orders) by default, manual promotion to live trading only after logged data supports it
- Localhost dashboard with bounds-checked, tighten-only parameter API and timestamped strategy backups
- GPS-based landmark audio tour app - React Native (Expo) frontend, Node.js/Express + PostGIS backend
- Offline content pipeline: Wikipedia/Wikidata facts -> Claude drafts sourced narration -> TTS synthesis, admin-reviewed before publish
- Live geofence triggers (250m radius) play narration automatically; every fact links back to its source
- Route scoring blends travel time, landmark count, and a crowd heuristic across 2-3 candidate routes
- Single Go binary merging a token-optimization pipeline with a policy/audit governance layer into one agent platform
- Autonomous agent loop with sandboxed code execution and web fetch, plus a web chat UI and terminal REPL
- Multi-provider LLM aggregation gateway behind one API endpoint - price/health routing, automatic failover, streaming translation, prepaid billing
- Atomic escrow billing, concurrency-tested (100 racing debits, zero overdraw); hash-chained tamper-evident audit log
- Turns a job posting into a personalized ~100-question interview study guide, cross-referenced against real interview discussion from public, terms-compliant community APIs
- Deliberately avoids sites that block automated access - treated as an authorization boundary, not just a scraping tradeoff
- Model-routed answers: a fast model for resume-grounded behavioral responses, a stronger model with capped web search for technical/system-design claims
- Mock mode by default (built-in question bank, no API key required); live mode degrades gracefully per data source
Research Output
Builds a real-time American Sign Language recognition system on single-board hardware so deaf and hard of hearing participants can join online meetings. Covers logistic regression, CNN training, training on microcontroller devices, and running the model in real time with Python, TensorFlow and Keras.
↗ View on ManningTechnical depth
Other projects
Side projects shipped for the developer community - open source and installable today.
Reads your Claude Code session logs and shows live token usage and cost estimates in your menu bar - broken down by today, this month, and all time. Per-model breakdown, Claude API status, auto-refresh.
cd ClaudeCosts && bash install.sh
Let's build
something smart.
Currently on expat assignment in the Netherlands through April 2027 - and actively exploring what's next. Open to senior OBD/diagnostics, EV systems, AI × automotive, and technical program management roles globally.
Everyday tools,
built right in.
10 web-based utilities you'll reach for daily - no sign-up, no ads, no tracking. Works entirely in your browser.
🧮 Calculator
Full scientific-style calculator. Keyboard-friendly.
📐 Unit Converter
Convert length, weight, temperature, speed, area, and volume.
🌍 World Timezones
Live clocks for major cities worldwide. Auto-updates every second.
📝 Word Counter
Paste or type text - get real-time word, character, sentence, and reading time stats.
🔐 Password Generator
Generate strong, random passwords instantly.
📷 QR Code Generator
Instantly generate a QR code for any URL, text, email, or phone number.
🎨 Color Picker & Converter
Pick a color, get HEX, RGB, HSL, and CSS variable instantly.
💻 Number Base Converter
Convert between Decimal, Binary, Hexadecimal, and Octal instantly. Essential for embedded/firmware work.
⏱️ Pomodoro Timer
The classic focus technique: 25 min work, 5 min break. Builds deep focus.
⚖️ BMI Calculator
Calculate Body Mass Index with metric or imperial units.