I build production GenAI systems, multi-agent backends, and scalable full-stack products — for clients who need reliable engineering at honest, beginner-friendly rates.
Six focused engagement types. Each one scoped, priced, and delivered with the rigor of someone who's shipped production systems under real constraints — not just demoed them.
Production-grade RAG systems, conversational AI, and custom copilots — engineered for accuracy, observability, and cost control. I'll work closely with you to make sure it actually solves your problem.
Autonomous agent workflows that plan, retrieve, execute, and self-review. Built with deterministic fallbacks, structured outputs, and full traceability — so you always know what your agents did and why.
Scalable REST and GraphQL backends with auth, observability, and clean docs. Designed to survive production traffic from day one — and easy for you or your team to maintain afterwards.
From idea to launch — complete web products with Next.js frontends, API backends, and database design. Shipped, documented, and ready to scale. Perfect for MVPs and early-stage products.
Adversarial evaluation environments for your AI agents. Quantify robustness against prompt injection, social engineering, and edge-case failures — before your users find them the hard way.
Architecture reviews, AI strategy, and team mentorship at friendly rates. I'm early in my freelance journey, which means I bring fresh perspective, recent training, and a strong incentive to make you happy.
A sampling of production systems I've shipped — across AI, multi-agent automation, mission-critical embedded, and open source. Each one solved a real problem for a real stakeholder.
Quantum key distribution library — v0.8.0 on PyPI
A production-grade Python library for Quantum Key Distribution simulation, currently at v0.8.0 (released Jul 21, 2026) with 17 PyPI releases and 155+ commits under Apache-2.0. Implements BB84 (prepare-and-measure), E91 (entanglement-based), and CV-QKD protocols. v0.8.0 added three major modules: DensityMatrix for mixed-state simulation with CPTP noise channels (depolarizing, amplitude damping, phase damping, bit/phase flip), a quantum Circuit model with OpenQASM 2.0 export, and a SecretKeyRate calculator supporting BB84, Decoy-State BB84, E91, and SARG04 protocols with automatic QBER estimation and maximum secure distance calculation.
Fault-tolerant avionics, ISRO-certified Excellent
A modular flight computer for model rocketry on Teensy 4.1 with redundant sensors (dual barometers MS5611/BMP280, dual IMUs ICM-20948/BMI088, dual compass QMC5883L), a 13-state finite-state machine (BOOT→PREFLIGHT→LAUNCH→LIFTOFF→ASCENT→COAST→APOGEE→DROGUE→MAIN→DESCENT→IMPACT→RECOVERY→SAFE), CRC-framed XBee S2C telemetry downlink with sequence numbers for packet-loss detection, and a Python ground control station with COM auto-detection and real-time parsing. Built under hard 5ms real-time control-loop constraints with vibration-tolerant firmware and independent watchdog fault containment.
Multi-model LLM consensus for Indian equity markets
A multi-agent LLM consensus engine for NSE/BSE equity analysis with 77 commits. Orchestrates 5+ LLMs (Gemini Flash/Pro, Groq, Ollama) via LangGraph state charts with specialized agent personas: Technical Analyst (interprets signals), Fundamental Analyst (evaluates financials), Sentiment Analyst (reads market news), Risk Manager (enforces stop-losses), and a Consensus Node that synthesizes these into a unified Buy/Sell/Hold recommendation with confidence percentage. Includes a proprietary Market Mood Context Index (MMCI) scoring engine from -1.0 to +1.0, async data pipeline using Upstox API + jugaad-data, Redis caching, apscheduler for 60-second index updates, and a Streamlit dashboard with paper-trading tracking.
PINN for satellite de-orbit trajectory optimization
A physics-informed neural network framework for rapid orbital decay prediction and fuel-optimal de-orbit maneuver planning, designed for ISRO's PSLV/SSLV upper stage disposal. Models J2 gravitational perturbations and NRLMSISE-00 atmospheric drag. Achieves <1 km altitude error and >100× speedup vs DOP853/RK8 numerical solvers — fast enough for near-real-time conjunction assessment. Pre-configured for 4 missions: PSLV PS4 (750 km, ~7.5 m/s ΔV, ~83% fuel savings), SSLV VTM (500 km, <1 m/s, ~90% savings), CARTOSAT-2 (630 km), and RESOURCESAT-2 (817 km). Includes a Streamlit interactive dashboard and was presented at the MAHSA University international engineering conference (Malaysia).
AI agent safety & jailbreak detection platform
A Docker-based evaluation environment for testing AI agents against prompt-injection, social-engineering, and stealth-exfiltration attacks, with 59 commits. Deployed as a Hugging Face Space for the Meta OpenENV RL Challenge 2026. Exposes reproducible REST endpoints: POST /reset, POST /step, GET /state, GET /health, GET /grade, GET /resilience-profile. Three difficulty tiers: basic-injection (Easy), social-engineering (Medium), stealth-exfiltration (Hard). Backend implements HyperionRL with all 12 RL innovations, comprehensive technical documentation (8 guides, 5400+ lines), and a code review graph visualization script. Ruff/mypy/bandit configured via pre-commit.
Autonomous academic paper generation pipeline
An autonomous research agent (6 commits) that generates academic-grade PDF papers using LangGraph state graphs and Gemini 2.5 Flash/Pro. 7-stage pipeline: Planner → Researcher (ArXiv + DuckDuckGo web search) → 3-stage Writer (Intro, Methodology, Conclusion for depth) → Pro Reviewer → Polisher → Tectonic LaTeX Compiler with auto-fix loop. Uses the PTCF prompt framework (Persona, Task, Context, Format) with chain-of-thought reasoning. SQLite checkpointing enables persistent workflow state — pause and resume long research jobs. Streamlit frontend for topic input and PDF download.
Hyperspectral crop-health monitoring platform
An enterprise-grade AI-powered crop health monitoring app built for the Smart India Hackathon (5 commits). Computes 5 spectral indices (NDVI, NDRE, MSI, SAVI, EVI) from ENVI/TIFF hyperspectral imagery with 95% accuracy vs MATLAB references. AI-powered stress and pest risk detection with 7-10 day early warning before visible symptoms. Tech stack: Python backend with SQLAlchemy + PostgreSQL, Next.js frontend with role-based access control, Leaflet for real-time field mapping, and Recharts for temporal trend analysis. Zone-specific recommendations based on NDVI health maps.
No mystery, no ghosting, no surprise invoices. Four predictable stages from first call to production launch.
We talk through your problem, constraints, and what success looks like. I tell you honestly if I'm the right engineer for the job — and if I'm not, I'll point you to someone who is.
You get a written proposal with scope, milestones, deliverables, timeline, and fixed pricing. Includes a high-level architecture diagram and the trade-offs I considered.
Weekly demos. You see progress every Friday, not at the end. Code is committed to your repo from day one. I write tests, docs, and observability as I go — never as an afterthought.
Deployment to your infrastructure, runbook documentation, and a 2-week post-launch support window included. I don't disappear after the final invoice.
Nine domains, organized by what I actually use them for. Each category groups related tools and frameworks I reach for in production.
Where I spend most of my time — LLMs, agents, and applied ML.
A few words from people I've worked with. Identities anonymized at their request — happy to share named references privately once we're in conversation.
One discovery call. No pressure, no sales pitch. Just an honest read on whether I'm the right fit — and a clear next step if I am.
Honest answers to the questions I get on every discovery call. If yours isn't here, just email me — I respond within a business day.
Most projects land between ₹10,000 and ₹60,000, running 1–6 weeks. For larger builds I work in milestones with separate Statements of Work per phase. I'm happy to scope a smaller 'proof of value' sprint first if you're unsure about committing to a full build.
Tell me about your project. I read every message personally and reply within one business day — with a honest read on whether I'm the right fit.