Who Thought Your Last Thought? The Space Between an Urge and an Action
A late-night message, a red apple, and a finger-and-clock experiment lead to a harder question: if a thought arrives before we can explain it, where does our response begin?
AI Reliability Engineering is the discipline of finding truth references. Founder of Qualixar — nine shipped open-source instruments, ten public arXiv preprints, and a daily dawn study of Vedanta and Sāṅkhya validation models.
How Vedanta and Sāṅkhya provide validation frameworks for understanding intelligence — artificial and embodied. Classical Indian philosophy was built for one goal: verifying truth references in dynamic, non-deterministic environments.
Balancing the three cosmic Guṇas (Tamas/inertia/regularization, Rajas/learning rate/momentum, and Sattva/convergence) to tune neural network hyperparameters.
Viewing static weights (Prakāśa) as latent potential, animated by dynamic runtime self-reflection loops (Vimarśa) and iterative token generation (Spanda).
Applying classical validation instruments (Pratyakṣa, Anumāna, Śabda) to build formal assertions and reliability verification checks for agentic runtimes.
Translating theoretical metaphysics into production code. Real-time contracts, assertions, and test pipelines.
Define YAML rules and check supplied structured agent state in Python. Place checks at an action boundary and raise on hard violations.
Store and recall project context across AI agent sessions through MCP or CLI, with workspace isolation, explicit operating modes and auditable retrieval.
Orchestration runtime of the Qualixar AI Reliability Engineering platform. 13 execution topologies (sequential, parallel, debate, mesh, hybrid, and more), Forge AI auto-design, adversarial judge pipeline, cost-quality-latency routing. 2,936 tests passing. Paper on arXiv.
Ten public preprints on arXiv. Formal methods, security, and memory for AI agent systems.
→ The instrument found 47 vacuous gates — checks satisfied by the absence of the thing they check — in shipped, reviewed code
Read on arXiv →→ Two instances of the same model failed together on 90% of missions where either failed — independence is not a safe assumption
Read on arXiv →→ Governed memory with scoped component evidence and disclosed negative results
Read on arXiv →→ One universal port for AI agents — 12 topologies, POMDP routing, four-tier degradation
Read on arXiv →→ Forgetting as a feature — biologically-inspired decay, cognitive quantization, multi-channel retrieval, zero LLM calls
Read on arXiv →→ 74.8% on LoCoMo with zero cloud calls — the highest local-first score reported
Read on arXiv →→ Regression testing that treats non-determinism as the subject, not the obstacle — 5–20x cost reduction
Read on arXiv →→ Agent skills execute with your agent's full privilege and nobody checks them — 675 tests, 8 novel contributions
Read on arXiv →→ Design-by-contract for agents — invariants that hold on every execution path. 1,980 sessions across 7 models
Read on arXiv →→ Shared memory is an attack surface — Bayesian trust scoring that degrades a poisoned writer instead of trusting it
Read on arXiv →Experience
Senior Manager & Solution Architect
Leading digital transformation programs for Fortune 500 European enterprises. Managing 100+ member teams across multiple concurrent projects. End-to-end solutioning with GenAI and Agentic AI.
5× Global Technology Innovation Award Winner (2021–2025). Built AI platforms for automated voice dubbing, sound effects generation, and video production — recognized in national media.
Software Engineer
Enterprise solution development for major US telecom and banking institutions.
Software Engineer
Digital platform development for global retail and telecom enterprises.
Network Engineer
Where the journey began — building networking infrastructure for telecom carriers.
Technical Expertise
Cloud & Infrastructure
Digital Experience
Commerce
Backend
Frontend
AI & ML
Architecture
Education
Dual Degree: B.Tech (Engineering) + LLB (Law)
Six decisions an AI solution architect actually owns. Each one is backed by something shipped and public rather than a slide — nine open-source instruments and ten arXiv preprints came out of making these calls and then having to defend them.
Which agent capability an organisation buys, builds, or refuses.
Vendor evaluation and onboarding, rollout topology, the landing zone agents run in, and the guardrails that make a pilot survivable in production. The hard call is usually what not to automate yet.
Evidence: Qualixar OS →What counts as proof that an agent is done and correct.
Behavioural contracts, independent gates the worker cannot write to, regression harnesses built for non-determinism, and drift detection. An agent asserting it finished is not evidence; a gate verdict in an append-only ledger is.
Evidence: Bounded Loops →What an agent is allowed to remember, and for how long.
Retrieval design, scoping and governance boundaries, retention and erasure, and the auditability that makes a memory store defensible rather than merely useful.
Evidence: SuperLocalMemory →What an agent is permitted to execute and with whose privilege.
Skill and plugin scanning, SBOM and attestation, least-privilege tool boundaries, and secret containment. Agent skills run with the agent’s full privilege, and most pipelines never check them.
Evidence: SkillFortify →Which model, tool, or human handles a given step.
Execution topologies, cost–quality–latency routing, and typed decisioning with calibrated confidence so routing is an inspectable record rather than a prompt-shaped guess.
Evidence: Jev Decision Layer →What evidence the organisation can produce when asked.
EU AI Act readiness, audit trails, evidence ledgers, and the operating model that keeps them current after the launch programme disbands.
Evidence: Research →Engagement details and client names are deliberately absent.
Daily study of Advaita Vedanta with Shankara’s commentaries. The tradition’s emphasis on pramāṇa — valid means of knowledge — directly informs the formal verification methodology used across Qualixar instruments. Vedanta is the original reliability engineering: a method for testing the mind’s outputs against an unchanging reference.
Tractable in code: every agent assertion is a prameya; every contract, regression, and supply-chain check is a pramāṇa; every production failure is a doṣadṛṣṭi — the cataloged flaw of perception that the system must be tuned away from.
Read the full essayI make films with my daughter about the questions AI made me ask. It’s a different room with different rules — spoken, not written. Enter the film library.
The AI Reliability Engineering Platform. Nine shipped instruments, ten public arXiv preprints, and the courses that teach how they were built.
A late-night message, a red apple, and a finger-and-clock experiment lead to a harder question: if a thought arrives before we can explain it, where does our response begin?
Across three centuries of famine and epidemic, women outlived men — strongest of all in newborns, before behaviour exists. A father and engineer traces the biology of endurance, and where it quietly meets Vedanta.
Everyone's asking if AI is waking up. I build AI for a living — and at 3 a.m., beside my four-year-old as his fever crossed 103, I understood we're asking the wrong question. The real divide isn't smart vs dumb. It's alive vs driven.
“The signal is always there.
You just have to build the filter.”
7 live instruments in the Qualixar suite. New research every few weeks.
Follow the build.