Where frontier research becomes product
TechNektar™ works across three fields that feed each other — AI, deep-tech engineering and fintech.
Three fields, cross-pollinating. Ideas from one sharpen the others.
Ancient epistemology tightens our AI. Aerospace-grade rigor tunes our control systems. Causal thinking from physics reshapes how businesses decide. We consult across all three.
Machines that reason
Mechanistic interpretability & AI safety, LLM fine-tuning, Claude Code tooling, creative world models, and foundation models built from scratch.
See AI research →Systems that run
Reinforcement-learned control of supercritical-CO₂ power cycles, digital twins & surrogate models, turbomachinery, and two decades of energy engineering.
See deep-tech →Decisions that deliver value
Causal inference when correlation lies, world models that plan in imagination, and the storytelling that makes analytics land in the boardroom.
See fintech →30+ open-source systems, papers and live apps — built to be used, not just cited.
01 · Artificial Intelligence
from understanding LLMs → to breaking them → to defending them
Active Circuit Discovery
An active-inference (POMDP) agent walks attribution graphs to find the features that causally drive a model's answer — on Gemma-2-2B and Llama-3.2, built on Anthropic's circuit-tracer.
Refusal as a Broken Symmetry
Refusal turns out to be a measurable, ablatable, dosable residual-stream direction — a shared necessary core across model families, linking jailbreak, hypnosis and vaśīkaraṇa.
Prabodha — the recognition-gated moat
A bring-your-model jailbreak-hardening platform. The activation-level moat cuts attack success as hard as brute-force hardening — at zero benign over-refusal.
Pratyabhijñā World Model
A Dreamer-class creative world model coupled to a frozen 120B LLM through a learned Vimarśa bridge. The world model imagines; the LLM speaks. 9 of 10 split hypotheses pass.
A foundation model, built from scratch
Pāṇinian Structured pretraining for small LAnguage Models: grammar-generated Sanskrit with gold parses, then real Sanskrit + English. Ranked #2 overall on the BabyLM 2026 strict-track leaderboard.
Sound as meaning
A speech-centred audio language model on a Sanskrit byte-core. The Sphoṭa-Lens localizes where meaning emerges (layer 13, 11× above chance) — and adaptation beats scaling.
Pratyabhijñā Creative Engine
Recursive self-reflexivity for LLM creative cognition — a Claude Code plugin that generates, judges and consolidates through a recognition cascade.
An AI music platform
Composition-aware instrumental and lyrical generation — a Next.js app orchestrating a DGX-hosted model fleet, end-to-end in ~39 seconds.
A machine-verified computation engine
Classical jyotiṣa as rigorous software: 196 passing tests, Lean 4 proofs with zero sorry, LLM narration verified against engine output — a template for verifiable domain engines.
TRIZ Engine
Systematic contradiction resolution with the 40 Inventive Principles — 327 tests passing, benchmarked on 4,900+ problems.
Pratyakṣa
Long-context discipline for Claude Code — Avacchedaka-typed retrieval, Khyātivāda hallucination taxonomy.
AttractorFlow
Steers multi-agent trajectories with dynamical-systems theory — Lyapunov exponents classify seven regimes and trigger interventions.
OpenClaw Swarm
Multi-agent Claude orchestration — role-based routing, Docker sandboxing, Telegram control, live telemetry.
02 · Deep-Tech Engineering
aerodynamics, predictive maintenance, supercritical-CO₂ power, avionics
Turbine blades, wind-tunnel proven
A cross-border collaboration with Politecnico di Milano put an advanced turbine cascade through a full wind-tunnel campaign — validating blade designs that became a revenue-generating product line.
Predicting turbine failure without failure data
VIGnAN — a GAN + LSTM system that synthesizes failure signatures to flag aero gas-turbine issues up to 500 operating hours early. First place, Dare to Dream 2.0 national innovation contest.
The gyroscope insight behind landmark missions
A years-long impasse in advanced gyroscope development broke when a previously unrecognized dynamic interaction was identified — new theory that flew on lunar and Mars missions.
Remote monitoring that pays for itself
A steam-turbine remote condition-monitoring platform — real-time health visibility, trend analysis and early warnings that cut unplanned downtime by up to 30% and lifted service renewals 40%.
Teaching an AI to run a power plant
Deep RL controls a supercritical-CO₂ Brayton cycle recovering waste heat from steel-furnace exhaust — trained in a physics-faithful OpenModelica digital twin, deployed at sub-millisecond latency.
Published research & a patent
- EOS-based analytical optimization of the sCO₂ Brayton cycle — J. Supercritical Fluids, 2021
- 10 MW recompression sCO₂ cycle for tropical climates — Applied Thermal Engineering, 2021
- Novel sCO₂ axial turbine design — ASME Turbo Expo, 2019
- Novel cycles for waste-heat recovery — Indian Patent, 2019
Decades in industry
Aero gas-turbine diagnostics & prognostics at General Electric; inertial-navigation avionics for launch vehicles at the Vikram Sarabhai Space Centre, Indian Space Research Organisation.
Postdoctoral research
sCO₂ turbomachinery for industrial waste-heat recovery at City, University of London; multi-institution European consortium leadership.
University research & teaching
Research and student supervision at the Indian Institute of Science and City St George's, University of London — bridging research and industry practice.
03 · Fintech & Causal Analytics
causal inference, world models, decision intelligenceDreamPrice — a causal pricing world model
DreamerV3 + Mamba-2, Hausman-IV causal identification and MOPO pessimism learn a retail-pricing policy entirely in imagination — from Dominick's historical scanner data.
Multivariate causal inference
Convergent Cross Mapping for time-series causality — who is really driving whom, when correlation lies. Correlation plots, forecasts, MAE/RMSE summaries.
The Coffee-Shop Mystery
Causal inference taught through a café: correlation traps, instrumental variables, double ML and transfer entropy — as Medium essays, YouTube episodes and an executable Jupyter Book.
Don’t take our word for it — run it, watch it, read it.
Every claim on this page opens into something you can use right now: live apps, research sites, storytelling episodes, invited talks and essays.
Medium ↗
- DreamPrice: An AI That Learns to Price by Dreaming
- The Coffee Shop Mystery — Part A: Enter Café Chaos causal-inference series
- The Coffee Shop Mystery — Part D: Flow Dance transfer entropy
- When the Context Window Is Big and the Agent Is Still Confused
- Pratyabhijñā Creative Engine
- Rivers of Thought: Beyond Generative AI
Substack ↗
Portfolio ↗
- Turbine blades, wind-tunnel proven aerodynamics · arXiv 2407.11210
- Predicting turbine failure without failure data predictive maintenance · contest winner
- The gyroscope insight behind landmark missions avionics
- Remote monitoring that pays for itself industrial IoT
Dense research, made to travel.
The same ideas, retold for different readers — a podcast, two newsletters, a case-study portfolio, and a machine-verified book.
TechNektar Newsletter
Demystifying AI, energy and engineering — the gateway essays. →
Long-form technical essays
"Teaching an AI to Run a Power Plant", the Coffee-Shop Mystery causality series. →
Case-study portfolio
Two decades of engineering wins — wind-tunnel campaigns, predictive-maintenance awards, avionics breakthroughs. →
Where philosophy meets architecture
The Proven Word
Vākya-Vallarī — a living, machine-verified edition of Bhartṛhari's Vākyapadīya: all 1,796 kārikās, each accepted reading proved against its contract by a Lean 4 kernel.
Bring us the hard problem.
Engagements from a two-day teardown to a multi-month build. We prototype fast, ship tested, and leave you the code and the reasoning. Problems that straddle two disciplines are a specialty — but a hard problem inside one is just as welcome.
AI — Foundation Models, Research Engineering & Safety
- Foundation-model research & from-scratch pretraining
- Domain fine-tuning & reasoning engines
- Mechanistic interpretability audits, jailbreak hardening & red-teaming
- Agent plugins, skills & MCP tooling for Claude Code
Deep-Tech Engineering
- Aerodynamics, turbomachinery & energy-cycle analysis
- RL & optimal control for industrial systems
- Digital twins & surrogate modeling
- Predictive maintenance & remote condition monitoring
Fintech & Causal Analytics
- Causal inference from observational data
- World models & decision intelligence
- Time-series and information-theoretic analytics
- Technical explainers & thought leadership
Every project above started as someone's "impossible" brief.
Bring us yours →Cross-pollinating innovation, literally.
TechNektar™ is an independent research & engineering practice, led by a data scientist and engineer with a PhD in supercritical-CO₂ power cycles and decades of frontier experience — from inertial-navigation avionics at the Vikram Sarabhai Space Centre, to diagnostics and prognostics at General Electric, to research at world-leading institutions.
The through-line is transfer: aerospace analytics informing demand models; 2,500-year-old logic tightening machine reasoning; reinforcement learning running a power plant. We publish in the open because tested ideas travel further — and we consult because the fastest route from idea to product is a team that has already shipped across all three fields. If your hardest problem sits between disciplines, it has come to the right place.
