AI & Machine Learning technologies
The AI stack,minus the hype.
Model choice matters less than retrieval quality, evaluation and cost control.
Most AI projects fail on retrieval and evaluation, not on model capability. We're deliberately model-agnostic — systems are built so a provider can be swapped as capability improves or prices move.
What we use
And why we picked each one.
Anthropic Claude
Frontier language model family.
WhyStrong reasoning and instruction-following, with enterprise terms that exclude your data from training.
OpenAI
GPT model family plus realtime voice.
WhyBroadest capability range and the most mature tooling ecosystem. Realtime API is currently the best option for voice.
Llama & Mistral
Open-weight models you can self-host.
WhyWhen data cannot leave your infrastructure. Weaker at hard reasoning than frontier models — we're direct about that trade-off.
LangChain & LlamaIndex
Frameworks for retrieval and agent orchestration.
WhyUseful scaffolding, though we drop to direct API calls where the abstraction costs more than it saves.
PyTorch
Deep learning framework.
WhyFor genuine model training — vision, forecasting, custom architectures. Where the problem isn't a language model.
pgvector & Pinecone
Vector stores for semantic retrieval.
Whypgvector first, to avoid a second database. Pinecone when scale genuinely requires it.
vLLM
High-throughput inference server for open models.
WhyMakes self-hosting economically viable at sustained volume. Below that, hosted APIs are cheaper.
LangSmith
Tracing and evaluation for LLM applications.
WhyEvaluation is the step teams skip and then regret. Without it you can't tell whether a change helped.
How we choose
Four criteria, in this order.
Fit
Does it solve your actual problem, or a more interesting one?
Hiring pool
Could you hire for it in India if you took the project in-house?
Total cost
Licences, hosting and the operational time to run it — not just build cost.
Exit risk
How hard would it be to leave? We don't build things you can't walk away from.
Services
What we build with it.
Usually several — cheap models for routine steps, strong models for hard reasoning. We build systems so the model is swappable, because the leader changes every few months and you shouldn't be rewriting each time.
Only if policy requires it or your volume is genuinely high. Below sustained heavy usage, hosted frontier models are cheaper, better and far less operational work.
Retrieval quality and missing evaluation. If the right context isn't retrieved, the model answers from memory and sounds confident doing it. Without evals, nobody notices until a customer does.