Krishna Kankipati

AI architect engineer product builder

AI should be accessible.
And reliable enough to use.

I’m Krishna. I work where large-scale systems engineering meets AI research, turning capable models into resilient, interpretable products people can actually use. That takes more than a good answer: it takes evidence, clear boundaries, and a way to see what happened.

Examine the systems Experience and complete project index
Working thesis

The interesting work happens between the model and the world.

Models interpret messy inputs. A dependable system also needs event-driven intake, source authority, explicit constraints, evaluation at the point of use, and a recovery path when work fails.

IngestInterpretConstrainDecideTrace
Selected systems

Four systems, different kinds of responsibility.

These are working systems and prototypes with consequential boundaries. Each dossier shows the problem, the constraint, the design move, and the artifact it produced.

Contract intelligence / production work

SpendRule

The answer has to survive the journey from inbound file to client decision.

Problem

Healthcare invoices and governing agreements arrive through different systems, often as scans or archives. Pricing authority may be split across amendments, rate tables, facilities, and effective dates.

Hard constraint

A financial conclusion needs the right source, scope, and time. An accepted upload event is not proof a document was processed; retries and replay must preserve identity and avoid duplicate decisions.

Architectural move

Route SFTP and dropzone files through event-driven archive and document handlers into a versioned vault. Run extraction, contract compilation, and invoice validation as durable stages. Use models for semantic matching, deterministic checks for pricing, and traces and audit paths to examine the result.

System shape From intake to accountable decision
SFTP feedDropzoneDirect upload
Archive / file events
Versioned document vault
Extraction + contract authority
Model match + deterministic check
Source-linked resultReview if unresolved

Retries, traces, and audit records help inspect the processing and decision path.

Evidence / artifact

A production validation flow with content-addressed document storage, event-driven intake and recovery, source-linked decisions, and an asynchronous audit of validation outcomes.

What this reveals

I design the operating and recovery path alongside the AI decision path. Provenance matters from the first file transfer to the final explanation.

Legal AI / platform engineering

Casey / 108-AI

Legal AI needs a workflow a reviewer can follow, not just an answer.

Problem

Contract review, clause drafting, comparison, and multi-document Q&A draw on different parts of a legal corpus. The work spans document processing, retrieval, model judgment, and a lawyer's review surface.

Hard constraint

Requests must stay within the right tenant and document context. Long-running review jobs need explicit state, retry and failure behavior, and a trace that shows how a recommendation was formed.

Architectural move

Separate ingestion and hybrid retrieval from LangGraph review, drafting, and validation stages. Wrap longer jobs in event-driven execution with status and failure handling; instrument graph nodes and model calls for inspection.

System shape Legal context to reviewable output
DocumentsPlaybooksRisk profiles
Chunking + hybrid retrieval
Review state graph
Draft + validate
Reviewable recommendationTrace + run status

Event steps and model traces give the workflow an inspectable history.

Evidence / artifact

Implemented contract-scoring and playbook-customization graphs, event-triggered workflows with failure handling, and document Q&A paths with grounding and citation checks.

What this reveals

I shape agent capabilities as bounded services with observable state, so product teams and reviewers can inspect the path to an output.

On-device safety / team project

SafeScreen AI

A visual safety layer that does not need to see your screen from a server.

Problem

Safety tools often ask people to upload the very screen content they want protected. An intervention also has to arrive while the content is still on screen.

Hard constraint

Screen frames stay on the phone. The Android app requests no internet permission, and a weak manipulation signal cannot become a definitive warning.

Architectural move

Run detection on-device, pass its output through an explicit policy layer, and blur across apps. Keep deepfake detection as a qualified secondary badge because its out-of-distribution evaluation was near chance.

System shape Privacy in the boundary
Screen frame
Local inference
Safety policy
Blur for explicit contentQualified badge for uncertainty

Device boundary: no internet permission in the app.

Evidence / artifact

Working Android screen monitor and blur overlay, validated on device; first place at the Qualcomm × Meta ExecuTorch Hackathon. Winners recap · Team entry

What this reveals

I prefer promises a user can verify in permissions and behavior, and I make weak model signals visibly uncertain.

Agent systems / team project

ARIA

An agent may suggest a trade. It should not grant itself permission to make one.

Problem

Autonomous trading agents can produce convincing rationales without making safe decisions. A trace is only useful if it also records what the system blocked.

Hard constraint

The agent operates in paper trading. Position, exposure, and rate limits are enforced in code rather than requested in a prompt.

Architectural move

Separate strategy reasoning from a risk router with code-enforced gates. Sign decision checkpoints and retain rationale and rejection records for later inspection.

System shape Reasoning cannot bypass the gate
Market signal
Agent proposal
Risk gates in code
Paper trade if allowedRejection if blocked

Both paths leave a signed decision record.

Evidence / artifact

A paper-trading system with signed checkpoints, ERC-8004 attestations, and rejection logs; second place in a hackathon prize track. Team entry · Source code

What this reveals

I separate what a model recommends from what the product permits, then preserve the evidence for both outcomes.

The full system inventory includes research work, document tools, interfaces, and smaller experiments.

From the field notebook

The evaluation unit changes the product.

Chronoscope is an active exploration in visual search across historical encyclopedia illustrations. Its key design decision is to judge the exact image crop or text chunk returned to a person, not a whole page that happens to contain something relevant.

Human labels remain ground truth; model judgments are review aids. That distinction is small in the interface and fundamental to whether a search-quality claim means anything.

See Chronoscope in the project index
Evaluation noteChronoscope / research
What was actually judged?
Old unitEntire source page
Working unitExact crop or OCR chunk
Ground truthHuman label tied to that evidence
An adjacent image is not proof that the retrieved image is relevant.
Operating principles

My default is to make the hard question visible.

Where did the answer come from?

Source identity, effective dates, and the exact cited artifact should travel with a conclusion. That is how a reviewer can challenge it.

SpendRule / Casey / Chronoscope

What is the model allowed to do?

Interpretation is useful. Financial rules, trade limits, and safety interventions need explicit policies that can be inspected and tested.

SpendRule / Casey / ARIA / SafeScreen

What happens when it is wrong?

Give uncertainty a real destination: human review, a qualified badge, or a blocked action with a recorded reason.

SpendRule / Casey / ARIA / SafeScreen

Can it recover and explain its own run?

Durable events, explicit status, retries, and traces turn stalled documents or failed model steps into work an operator can diagnose and resume.

SpendRule / Casey
About me

I build where AI research meets real-world systems.

My work has never stopped at building a model. I care about the engineering that makes AI dependable: source data, product decisions, evaluation, recovery, and an explanation someone can inspect. That is how advanced AI becomes useful beyond the people who built it.

I'm a founding AI engineer at SpendRule, working on healthcare contract intelligence. Before that, I built legal AI workflows at 108-AI / Casey Legal and worked across investment research, public-sector search, and computer vision. Across those domains, the question stays the same: can people trust this system when it meets the real world?

Read the full recordGitHubLinkedIn