I'm Pranay M Mahendrakar — an AI specialist working across large language models, natural language processing and computer vision. I've delivered 250+ AI systems spanning the automotive, defence and education sectors, and I write about the theory underneath them.
Applied AI for industry and defence, production platforms serving live businesses, and open-source tooling for the research community.
A multilingual, fully offline AI avatar built to handle showroom sales conversations without depending on cloud inference — engineered for privacy, latency and reliability on the floor.
Real-time computer vision for autonomous aerial security — detecting and classifying threats from live drone feeds under field conditions where false positives carry real cost.
Automation platforms that compress the grunt work of academic research — literature handling, analysis pipelines and reporting — deployed across two of India's leading institutes.
A voice assistant running a custom fine-tuned model entirely on-device — no network round trip, no data leaving the machine, and latency low enough to feel conversational.
A financial diagnostic platform that scores a company across eight vital signs — profitability, liquidity, turnover, risk — then points at exactly where margin is leaking and what a fix is worth.
A self-hosted assistant with retrieval-augmented generation over uploaded documents, a live terminal, multi-channel inbox and broadcast — wired through real-time WebSocket streams.
A personal-management mobile product backed by a hand-rolled REST API — registration, contact graph and identity flows built to run reliably on modest infrastructure.
An operations backbone for an EV manufacturer — invoicing, production tracking, dealer management and field sales consolidated into one authenticated workspace.
A programming language designed and implemented from the ground up — grammar, parser and runtime — as an exercise in taking a language from specification to something that actually runs.
A custom fine-tuned large language model packaged and deployed on Ollama for local inference — part of ongoing work on making capable models run without a datacentre behind them.
A multi-agent research pipeline — search, summarise, cite and fact-check agents working in concert — running entirely in the browser with live streaming and score visualisation.
An ongoing body of public tooling: an arXiv paper analyser, an LLM benchmarking framework, a self-evolving model ecosystem, and a suite of MCP plugins for Claude and ChatGPT.
Work on the semantics, limits and failure modes of large language models — interpretability, alignment decay, multilingual hallucination, formal verification and low-resource language equity. Every paper carries a permanent DOI and is free to read.
A survey of the circuits known to underpin in-context learning — still the most striking and least understood capability of large language models — and the problems left open.
Disambiguates four separate things the field calls "memory", and maps the long-context reasoning frontier now that state-space models have moved into production.
A measurement framework for capability degradation that compounds round over round — the alignment tax studied across iterations rather than in a single comparison.
Do negotiating model agents spontaneously develop channels human observers cannot decode? A conceptual framework and an experimental protocol for finding out.
Post-hoc detection and watermarking are routinely conflated. Separating them shows what has actually been solved, what has not, and where bias enters the picture.
Three distinct gaps between verifying robustness on ResNet-scale networks and verifying safety properties that matter, plus a specification-first research agenda.
A typology-aware agenda comparing hallucination across Dravidian and Indo-Aryan languages, where aggregate multilingual benchmarks flatten real structural differences.
Over 300 sign languages serve some 70 million Deaf signers, yet research overwhelmingly targets ASL. A research agenda for Indian and African contexts.
Why the barrier to spiking networks on neuromorphic hardware is a three-layer integration gap rather than the algorithm layer everyone optimises.
A critical assessment of whether energy-based models are a genuine alternative to autoregressive reasoning trained with reinforcement learning — and where the evidence stops.
Identifiability assumptions in causal representation learning, and the distance between the synthetic settings where they hold and the real video where they are needed.
A unified categorical framework for semantic coherence — treating meaning as a sheaf and locating hallucination in the failure to glue local sections into a global one.
Introduces the Quantum Mirror Framework, on breaking the barriers between human consciousness and artificial intelligence.
The research foundation under the deployed drone threat-detection work — proactive aerial surveillance architecture for security operations.
On the gap between models that convincingly generate emotionally appropriate responses and models that actually process emotion — and why the two keep getting conflated.
ORCID: 0009-0003-7224-029X
Three published books on artificial intelligence, and registered intellectual property in blockchain infrastructure and social-network integrity.
An accessible route into artificial intelligence for readers coming in cold.
A wider survey of the field and the directions it is pulling in at once.
Large language models explained from the inside out — mechanism before mystique.
A decentralised architecture for distributing cloud compute across untrusted nodes.
Detecting duplicated and impersonating identities across social platforms.
A show about the part of company-building nobody ships a framework for — the operator underneath the operation.
Before building a startup, build the person capable of building one.
Latest episode — Building Your Life's Mission
Episode 12 · 14 July 2026
Most people in AI pick a side. Either you write the papers, or you ship the systems. I've never found a good reason to choose — the theory gets sharper when something has to survive contact with a factory floor, and the systems get better when someone has actually read the literature.
So the work runs on both tracks. On one side: papers on sheaf-theoretic semantics, hallucination and the limits of machine reasoning, three published books, and registered patents. On the other: offline AI avatars for Mercedes-Benz, threat detection for the Indian Army, and research platforms serving thousands of users at IIT Bombay and IISc.
I work as Nodal Coordinator at IIRS-ISRO and teach as an Instructor at Tutorials Point, alongside running Sonytech. A strong bias runs through all of it: models should run where the data already is — on-device, offline, under your own control.
Model work rests on data structures and algorithms, so I keep that edge sharp deliberately — 3,235 problems solved on LeetCode, weighted toward the hard end.
problems solved of 4,028
across every difficulty tier
Code, courses, papers and packages — the work is public and verifiable.
Open to research collaboration, applied AI engagements, speaking and teaching. The fastest way to reach me is email.