Santosh Jammi Enterprise AI architect — home

Visakhapatnam, IN · UTC+05:30

Selected work

04 records

What I built, what stage it reached, what it cannot do yet

I describe my own projects the way I would want a vendor to describe theirs. Every entry states the problem, my actual role, the decisions that shaped it, the maturity, the evidence, and the limits. Nothing here is presented as a client success story, and no client names appear.

How to read the labels. Architecture means a designed system with documented boundaries, not running software. Prototype means working code proving the hard parts, on test or synthetic data. Working system means I use it for real work, but it is not packaged or multi-tenant. Pilot means used by someone other than me under observation. Production means operating for real users with support obligations. Nothing on this page is a production deployment at a client.
Architecture Prototype Working system
Record 01 2026

Enterprise agentic engineering platform

Governed AI engineering — scoped missions, evidence gates, human approval

Working systeminternal use Role: Sole architect and operator

The problem

Coding agents can produce plausible changes quickly, but the tooling around them rarely enforces that a change was scoped, reviewed, tested or reversible. Teams end up with two bad options: block the agents, or accept output nobody verified. The missing piece is not a better model — it is an engineering lifecycle the agent has to operate inside.

What I did

I designed the operating model and run it daily. A mission arrives as a frozen brief: the problem, the boundaries it may not cross, the acceptance criteria, and the stop conditions. Implementation is delegated; verification is a separate step with its own evidence. Anything consequential waits for my approval.

Decisions that shaped it

  • Split deterministic from probabilistic responsibilityBuilds, tests, linting, diff policy and evidence capture are performed by ordinary software. Models plan, write and review. Neither is allowed to do the other's job, so correctness never depends on a model's mood.
  • Freeze scope before dispatchAn agent gets one self-contained task with explicit boundaries. Vague scope is the main cause of confident, wrong work — and of merge conflicts when several agents run in parallel.
  • Separate implementation from verificationThe agent that wrote the change does not get to declare it correct. A verification pass re-reads the artifact, runs the checks, and records raw output. A sub-agent's summary is treated as a claim, not a result.
  • Keep the human at the gateDeployment, publishing, credentials and destructive operations are approval events, not agent decisions. Autonomy stops where consequences start.
Record 02 2026

DPDP compliance discovery & intelligence

Finding personal data, and proving why it is lawful to hold it

Prototypevalidated components Role: Product architect

The problem

India's Digital Personal Data Protection framework asks organisations to know where personal data lives, why it is held, how consent was obtained, how long it is kept, and to whom it has been disclosed. Most organisations answer with questionnaires filled in from memory. The evidence, when it exists at all, is scattered across shares, databases, spreadsheets and scanned forms.

What I did

I architected the platform end to end: asset discovery, classification, an evidence model, obligation mapping, risk scoring and gap analysis. I also built the validation harness used to check whether personal-data signals are being detected correctly, and defined what a reviewer needs to re-verify a finding.

Decisions that shaped it

  • Rules stay authoritativeObligations, retention limits and disclosure rules are encoded deterministically, because a compliance answer must be the same tomorrow. AI is introduced only where the work is genuinely ambiguous: classification, entity resolution, synthesis of findings.
  • Every finding carries its provenanceA finding without a source location, an extraction method and a confidence is not a finding — it is an opinion. Reviewers can open the source and disagree with the system.
  • Separate signal from conclusionDetecting a pattern that looks like a PAN or an Aadhaar number is a signal. Deciding that a system is non-compliant is a conclusion that combines signals, context and policy. Keeping them apart stops precision problems from becoming legal problems.
Record 03 2026

Local-first AI platform & model governance

Measure the model, then decide — and keep the GPU honest

Working systemlocal environment Role: Architect and operator

The problem

Local models are easy to start and hard to operate. Throughput, context limits, memory residency and quality differ per model and per runtime, and a single GPU gets oversubscribed when language, image and video workloads compete for it. Without measurement, 'we run models locally' means 'we hope'.

What I did

I designed and operate a local AI environment: model qualification, runtime comparison, residency and lease tracking, workload monitoring, and isolation between LLM inference and generative-media jobs. Sensitive work stays on the machine by default; anything that leaves it needs a documented reason.

Decisions that shaped it

  • Qualify, do not adoptEach candidate model/runtime pair is measured for throughput, memory behaviour under concurrency and effective context before it is used for anything that matters. Benchmarks from other people's hardware are treated as hints.
  • Residency is a resource with a leaseA loaded model occupies memory whether or not it is answering. Tracking residency explicitly, with reconciliation against actual processes, is what stops a 'stopped' model from starving the next job.
  • Isolate competing GPU workloadsLanguage inference and media generation have different memory profiles. Separating them keeps one long render from taking down an interactive session.
Record 04 2026

Restaurant operational intelligence

Why margin moved, which recipe version did it, and where to look next

Architectureprototype components Role: Product architect

The problem

Restaurant operators are shown what happened — sales, costs, variance — but not why. When food cost moves, the cause is usually a recipe change, a supplier price, a yield assumption or waste, and the dashboards available cannot tell them apart. So the operator investigates manually, or does not investigate at all.

What I did

I designed the intelligence layer over inventory, recipes, purchasing, suppliers, consumption, sales and profitability: the data model, the time-window analysis, expected-versus-actual consumption, and historical recipe resolution, with an output aimed at decisions rather than charts.

Decisions that shaped it

  • Recipes are versioned, effective-dated entitiesA dish's cost depends on which recipe was in force on the date of sale. Treating a recipe as a single current record silently corrupts every historical comparison, so recipe versions carry valid-from and valid-to dates.
  • Compute expectations deterministicallyExpected consumption is arithmetic: units sold times recipe quantities, adjusted for yield and unit conversion. Variance is then attributed to price, portion, waste or mix — with the arithmetic shown, not hidden behind a score.
  • Answer the operator's next questionThe product states what changed, what it cost, and the two or three places worth investigating. A model may phrase the explanation; it does not produce the number.
05 In progress

Also in progress

Earlier-stage work that is real but not yet worth a full case study.

GENESIS → PROMETHEUS AI media pipeline

A multi-stage production pipeline separating story, dialogue, character continuity, shot planning, performance intent, voice binding, image generation, animation, audio and rendering. Frozen production contracts and validation gates catch character drift, dialogue loss and voice mismatch before render — because a defect found after rendering costs hours, not seconds.
Architecture · validated runs

AI-driven news & content platform

A local-LLM-first publishing workflow built around a canonical 6Ws + H structure, with evidence handling, structured planning, generation, search and answer-engine checks, provenance, and publication-readiness gates. Language is generated; facts, policy, structure and validation stay deterministic.
Architecture · internal use

Auction intelligence

Source discovery, acquisition, normalisation, structured storage and document intelligence for fragmented auction data — with source reliability and provenance established before any AI interpretation is allowed to touch a record.
Architecture · prototype