01evaluations
Evidence over adjectives
“It feels better” is not a result. Every change ships with a measured delta against a baseline both sides agreed on before the work started.
AboutEngineers first.
Engineers first.
Consultancy second.
Hawthorn AI Labs is a small, senior team that builds the machinery between a model and the real world — the context it receives, the tools it can touch, the evidence it leaves, and the lessons it is allowed to keep.
05Hawthorn AI Labs
about --format json
$focus: custom models · agents · deployment
$model: senior engineers, no pyramid
$every engagement: eval baseline first
$handover: code, weights, runbooks, you name it
india · working worldwide
- Founded
- 2025
- Based
- India
- Reach
- Worldwide
- Open source
- Public repos
01 / Position
The future isn't just bigger models.
It's smaller, sharper systems around them — the retrieval, the tool contracts, the policy gates and the evaluation loop that make a model trustworthy in one specific job.
Most organisations we meet have the same experience: an impressive demo, a growing pile of prototypes, and no reliable path from either to production. The model was never the hard part. The hard part is everything the model needs in order to be safe, measurable and affordable at volume.
That is the whole company. We take ownership of the data contract, the model, the agent runtime and the deployment target, and we do not consider the work finished until your team can run and extend it without us.
6 wksmedian time to first production slice
100%of engagements start with a baseline
02 / Principles
Four rules we don't bend.
02cost & latency
Smallest model that clears the bar
The largest available model is rarely the right answer. We size the model to the requirement, then make that model genuinely good at the task.
03governance
Autonomy is a dial
Where an agent may act without a human is a decision we make explicitly, per action class, in writing — and enforce outside the model.
04ownership
Handover is part of the deliverable
Architecture decisions, runbooks, eval harness and a recorded walkthrough. You should never need us to keep the system alive.
03 / Track record
Talk to an engineerBuilt by people who have shipped.
Frontier-lab evaluation work, production anomaly detection at national scale, and open-source systems other builders rely on.
Custom models & agentic systems for enterprises
Building the data, model and deployment layers that turn a promising prototype into infrastructure a business can depend on.
Multilingual voice evaluation & training
Evaluation design and training data work for Hindi and English voice intelligence at frontier-lab scale — where a bad rubric silently ruins a model.
142K-record anomaly detection, in production
A multi-agent system across 600–700 collection centres. Two of three randomly audited, AI-flagged centres were confirmed in a secret audit.
Open source & community
Published agent, memory and governance systems used by other builders — plus a 100K+ audience, a podcast with 2M+ listens, and large-scale hackathons.
04 / Open source
View GitHubWe build in public.
The patterns above ship as reusable systems. Read the source before you hire us.
Learning governor
SkillLoop
Local-first harness that turns agent traces into governed evaluations, memories, skills and fine-tuning exports.
PythonEvalsMemory
Agent runtime
Governed Agent Architecture
Orchestration with durable memory, hybrid retrieval, injection defence and evidence-first execution.
LangGraphRAGGovernance
Persistent memory
SuperMem
Four-tier retrieval across full-text search, a knowledge graph, vectors and an LLM agent — exposed over MCP.
MCPFTS5Graph
Memory without RAG
Recall
An agent-navigated local knowledge base for persistent memory across desktop assistants.
Local-firstMCPPython
Small senior teams
The people designing the system are the people writing it. No handoff to a junior pod.
Direct access
You talk to the engineer doing the work, in your timezone overlap, without a account manager relay.
Long-term minded
We would rather ship something modest you can operate than something impressive you cannot.
Work with usTell us what
Tell us what
should exist.
If it can be specified and measured, it can be built. If it can't yet, that is exactly what a diagnostic sprint is for.
Typical first response within 24 hours · NDA-friendly