Shrinking a 16B Model: Distill, Quantize, Prune
Distillation, quantization and pruning compose - but only if you measure degradation on tasks, not perplexity. How we shrink bb-core 16B in-house.
Nithish Rajan8 July 202611 minAI EngineeringServing LLMs at Scale: KV-Cache to Quantization
KV-cache math, paged attention, continuous batching, quantization and TP: how we engineer LLM serving for throughput, cost-per-token and tail latency.
Magesh Sundar24 June 202611 minResearchPPO vs DPO vs GRPO: Preference Optimization Map
PPO, DPO, GRPO: a practical, gradient-level map of preference optimization for reasoning models - and how to actually choose.
Nithish Rajan10 June 202612 minAI EngineeringEngineering a 256K Context Window: RoPE & FlashAttention
A dense 16.4B decoder with a real 256K window. The positional-encoding math, IO-aware attention and honest evals behind it.
Magesh Sundar27 May 202611 minAI EngineeringDense vs MoE at 16B: The Architecture Trade-Off
Why decoupling capacity from FLOPs is seductive on paper and a serving liability on owned hardware - and how that math made bb-core a dense 16.4B decoder.
Magesh Sundar13 May 20269 minAI Engineeringbb-core 16B: Pretraining a 16B-Parameter Model
A field report on pretraining bb-core 16B: 16.4B dense parameters, 256K context, ~3.1T tokens, 2M+ GPU-hours, running on infrastructure the client owns.
Magesh Sundar29 April 20269 minApplied AIAgentic AI & MCP: Building Agents You Can Trust
Most agent demos die in production. Here is an honest architecture for autonomous systems that plan, act, verify and stay grounded in your source of truth.
Nithish Rajan15 April 20269 minApplied AIProduction RAG That Doesn't Hallucinate
A demo RAG impresses in a meeting. A production RAG survives an audit. Here is the engineering that separates them.
Magesh Sundar25 March 20269 minResearchBudget-Forced Reasoning: Frontier Accuracy, Less Compute
Force a token budget during reasoning training and the model learns to think efficiently instead of verbosely. Here is the mechanism, and why we own it.
Magesh Sundar4 March 20269 minResearchBeacon SI 2.5: Beating GPT-4o on Risk Detection
We trained a small reasoning model that beats GPT-4o on suicidal-risk detection. Here is the eval-first, SFT to DPO methodology behind it.
Nithish Rajan11 February 20269 minAI EngineeringTraining LLMs From Scratch vs Fine-Tuning: Our Thesis
Fine-tuning rents someone else's priors. Here is where that breaks structurally - and why we built bb-core 16B from the tokeniser up.
Nithish Rajan21 January 20269 minGuidesChoosing an AI Development Company in India
A build-vs-rent field guide for founders and execs choosing an AI development company in India - the framework, the questions, and the red flags.
Nithish Rajan9 January 20268 minReading this because you're building something?
If any of this maps to a problem you have, tell us. We engineer AI from scratch when that's what the problem needs - and say so when it isn't.