I'm Aakashdeep Srivastava, an AI Engineer with 5+ years taking ML systems from research → production: LLM orchestration, real-time computer vision, and high-throughput Python microservices. I don't stop at a notebook — I ship sub-40ms inference on serverless GPU, with reproducible Poetry environments, A/B deployment gates, and drift monitoring.

Looking for: AI / ML Engineer · LLM & Agentic Systems · Computer Vision · MLOps — full-time, end-to-end ownership.

• 🏗️ Chief AI Engineer @ Reclevo Infotech — building AI + logistics infra for circular-economy waste systems
• 🎓 B.S. Data Science & Mechanical Engineering @ IIT Madras — CGPA 8.9/10
• 📜 Azure Data Scientist Associate (DP-100) · Microsoft Data Analyst (PL-300) · MLOps Specialization
• 🏆 5× hackathon winner · ranked 2nd / 900 teams at Karnataka State Police Datathon
• 🌐 Portfolio & writing → skydeep.in

Specialised agents. Shared memory. Human in the loop. I build agentic systems where an orchestrator plans, routes, and verifies across tool-using agents backed by RAG — and a human stays in the approval loop. In production this cut Level-1 support tickets by 60% (OpenAI + Gemini agentic workflows at Reclevo).

End-to-end AI + logistics platform for rapid waste disposal.

• Waste-segregation ML pipeline: YOLO11-Seg + TensorRT INT8 → 84% field accuracy at <40ms, SAM 2 auto-labeling, Moondream2 VLM fallback for ambiguous classes. Deployed on Modal serverless GPU, 400+ images/day.
• Shipped Reclevo — a Rapid Waste Disposal Platform serving 5 user roles across multiple React Native frontends on a GoLang (Gin) microservices backend at 99.9% uptime.
• Optimised fleet logistics (SWAM) with data-driven routing → −25% fuel, faster pickup turnaround.
• Built a Pandas + NumPy ETL pipeline over GPS telemetry (route optimization, heatmaps, anomaly detection).
• Agentic workflows (OpenAI + Gemini) for support & waste categorization → −60% L1 tickets; Power Automate compliance flows → −95% manual entry errors.

MLOps · PEFT · FastAPI · AWS

• PEFT/QLoRA fine-tuning of Qwen-3B on SageMaker — 4-bit quant + gradient checkpointing + Optuna sweeps → 95% of baseline MMLU at −60% GPU memory vs full fine-tuning.
• Built an MLflow evaluation pipeline with A/B deployment gates and Poetry-standardized, reproducible envs → experiment tracking across 100+ models.
• Architected a FastAPI + Celery video-generation pipeline (Gemini Pro + ElevenLabs + Rhubarb Lip Sync), Redis broker, 50+ concurrent requests at 98% viseme accuracy, Docker + Prometheus health checks.
• Automated exam-notification ETL (SmartScraper AI + Pandas) → 1000+ notifications/day at 99% accuracy, +40% engagement.

Next.js 15 · NestJS 11 · PostgreSQL 16 + PostGIS · Python FastAPI · Anthropic API AI-powered election intelligence for Goa MLAs: booth-level analysis from public Form 20, an AI Co-Pilot with daily confidence-scored, evidence-backed recommendations, an OR-Tools visit optimizer, and a grievance CRM. Multi-tenant with RLS forced, 6 BullMQ worker pipelines (z-score/EWMA early-warning, Apify social scanning), OCR ingest via Sarvam Vision — all behind a PII firewall with prompt-injection defense. Deterministic SQL scoring; the LLM only writes explanations, never numbers.

Real-time sign-language recognition (CNN image classification + object detection, 95% accuracy) with emotion detection for autism support, plus speech-to-image generation (500+ voice commands, 90% satisfaction). Node.js/Express + WebSocket backend handling 100+ concurrent video streams at <100ms latency.

DPDP-Act-2023-compliant, zero foreign LLM calls. Three modules: Sathi (voice-first registration via Sarvam ASR/TTS), VargBot (TF-IDF + Sarvam-30B classifier, 14 ONDC domains, 98.5% accuracy), JodakAI (multi-factor scoring, NDCG@3 0.879). Next.js 15 + FastAPI + PostgreSQL 16 + Redis, full DVC pipeline with model registry + drift monitoring and an immutable audit trail.

PowerBI analytics + Redis async queue over 50+ police stations → −60% resource-allocation time, −40% response time; Airtable real-time tracking of 1000+ daily updates, −80% manual entry, 5-minute decision cycles.

Languages / OOP     Python (async, dataclasses) · GoLang · SQL · TypeScript · Pandas · NumPy · Polars
LLM & Agentic AI    LangGraph · LangChain · Google ADK · Gemini / OpenAI / Anthropic · RAG · ChromaDB · Pinecone
ML / DL / CV        PyTorch · TensorFlow · YOLO11-Seg · TensorRT · SAM 2 · Moondream2 VLM · PEFT / QLoRA · OpenCV
Pipelines           FastAPI · Flask · Celery · Redis · Kafka · WebSocket · Poetry · Docker Compose · CI/CD
Cloud & MLOps       AWS (SageMaker / Lambda / S3 / EC2) · Azure ML / App Service / DevOps · Modal · MLflow · DVC · W&B

• Indian Institute of Technology, Madras — B.S. Data Science Applications & Mechanical Engineering · CGPA 8.9/10
• Azure Data Scientist Associate (DP-100) · Microsoft Data Analyst (PL-300) · Machine Learning Operations Specialization
• Super Coordinator, Technicals @ IIT Madras — secured ₹5 Lakhs sponsorships across 18 hackathons, led a team of 6.

🥇 Product Space GenAI Hackathon (GTM) · 🥇 Sustainability Hackathon GIM'25 · 🥇 Education Equity Hackathon HackerEarth'26 · 🥈 Karnataka State Police Datathon (2nd / 900) · 🏆 Certificate of Excellence, ML Challenge 2.0 (IIT Madras)

_{☕ Built something useful from my work? You can sponsor a coffee — entirely optional. The Sponsor button up top works too.}

Open to: AI / ML Engineer roles · LLM & agentic systems · computer vision · MLOps — end-to-end ownership.

_{⚙️ I build systems, not just features. · Goa, India · skydeep.in}