case backlog cleared
On-premises compliance review workstream in a healthcare payer environment.
Senior Applied AI EngineerHealthcare · Insurance
Shailesh Dudala
I build AI systems that survive real operations.
From claim packets and clinical documents to HL7/FHIR events and human review queues, I turn messy, regulated data into validated, traceable workflows.
Agentic AI · Document Intelligence · Healthcare & Insurance · ML Platforms
System flight recorder
trace.claim_packet.042Synthetic artifact · no private data
{ "doc_type": "EOB", "pages": 9, "contract": "claim_fact.v3" }- 01Ingestpacket accepted
- 02Classifypages classified
- 03Extractfacts extracted
- 04Validatedate rule failed
- 05Human reviewreviewer routed
- 06Observetrace persisted
Reading mode
Selected outcomes across healthcare, insurance, and analytics delivery
Résumé-supported · scoped by contextreview-time reduction
Measured document-review workflow after local retrieval and structured extraction were introduced.
Approximate reduction in the measured workflow; not an enterprise-wide claim.automated closure improvement
Healthcare quality-measure evidence extraction and review workflow.
FWA waste reduction
Transportation anomaly detection and explainable review prioritization.
client P4P impact
Healthcare analytics programs supporting quality and value-based care delivery.
Approximate client performance-based payouts supported by the broader program.Selected work
Systems where the hard part begins after the model responds.
These stories focus on contracts, validation, review paths, and operational controls—not a gallery of model demos.
From claim packet to auditable action
A governed document workflow that turns mixed claim packets into typed facts, visible exceptions, and reviewer-ready decisions.
Reduced document handling in a measured workstream while keeping every uncertain field and route inspectable.resume-supported · Role, workflow scope, and qualified outcomes are supported by the latest résumé.
Read case study {
"packet_id": "SYN-042",
"contract": "claim_fact.v3",
"confidence": 0.78,
"rule": "service_date_sequence",
"state": "review_required",
"trace_id": "tr_8a21"
}- Ingest
- Classify
- Extract
- Validate
- Human review
- Action
- Observe
Clearing a 7,000-case review backlog without moving regulated documents off-prem
A local document-review system combining OCR, retrieval, small language models, and a reviewer surface inside the data boundary.
Cleared a 7K-case backlog and reduced review time by about 90% in the measured workflow.resume-supported · Backlog and review-time outcomes are supported by the latest résumé.
Read case study review_case:
source: synthetic-policy-17.pdf
pages: [4, 5]
retrieval_score: 0.84
extraction_state: qualified
route: reviewer_confirm
reason: conflicting_effective_dates- Local intake
- OCR
- Chunk
- Retrieve
- Generate
- Review
- Measure
Standards-aware patient communication, built as a team
A healthcare interoperability project recognized by the Global HL7 AI Challenge for Transformative Impact in Healthcare.
Team recipient of the 2025 Global HL7 AI Challenge recognition for Let’s Talk Doc.public-source · HL7 confirms Let’s Talk Doc and the Transformative Impact in Healthcare award.
Read case study {
"resourceType": "CommunicationRequest",
"status": "active",
"subject": { "reference": "Patient/SYNTHETIC" },
"payload": [{ "contentString": "Post-visit summary" }],
"language": "es-US"
}- Clinical context
- FHIR contract
- Patient-friendly transform
- Safety checks
- Human touchpoint
Turning model-behavior research into a reproducible engineering workbench
A public local-first workbench for activation steering, model inspection, API-driven experiments, and repeatable evaluation.
Public code, tests, UI, and documented limitations make the research inspectable rather than promotional.public-source · Source, tests, UI, and documentation are public.
Read case study experiment: sentiment-axis-04
model: local-transformer
layer: 18
coefficient: 0.65
seed: 42
comparison: baseline_vs_steered
status: reproducible- Load model
- Capture activations
- Construct vector
- Apply hook
- Evaluate
- Compare
How I build
Five habits that keep uncertainty visible.
The same operating principles connect claims, clinical documents, predictive ML, and local-model research.
- 01
Contracts before prompts
Typed inputs and outputs expose ambiguity before it reaches downstream work.
Claims intelligence ↗ - 02
Evaluation before automation
Define failure modes, acceptance sets, and release gates before scaling volume.
LLM Steering Lab ↗ - 03
Humans where risk remains
Route uncertainty with a reason and evidence instead of hiding it behind confidence.
Reviewer boundary ↗ - 04
Local-first when data requires it
Choose deployment boundaries around privacy, control, and operational ownership.
On-prem review ↗ - 05
Observability is part of the product
Trace latency, cost, tool events, errors, routes, and outcomes from the start.
Flight recorder ↗
Experience
Increasing scope, one throughline.
Biomedical data became healthcare prediction, then analytics products, on-prem GenAI, and agentic insurance workflows.
2026 — present
Insurance claims and agentic AI
MetLife via BizintexDesigning governed document and claims workflows with typed contracts, validation gates, reviewer fallbacks, and trace telemetry.
- ≈90% lower document-handling effort in a measured claim-packet workflow
- ≈50% shorter time-to-claim-payable in a supported workstream
2023 — 2025
Payer AI modernization
Inland Empire Health Plan via InfowaveLed local RAG/OCR, healthcare quality evidence extraction, predictive ML, FWA analytics, MLOps, and operational reporting.
- Cleared a 7K-case review backlog with ≈90% lower review time
- Improved automated measure closures by 20%
2020 — 2023
0-to-1 healthcare analytics platform
HexploraBuilt and scaled a predictive analytics platform across nine healthcare programs, combining risk models, data products, and care-manager workflows.
- $500K in new revenue supported
- ≈$3M in client performance-based payouts
Earlier foundations
Research and operational foundations
CommonSpirit Health · Health New England · University of Chicago · AbbVieWorked across hospital analytics, provider data, biomedical research, clinical sensors, genomics, and public-health modeling.
- Analytics spanning 142 hospitals
- 100K+ provider records validated
Recognition
Team credit, project names, and evidence stay attached.
Awards are supporting context—not a substitute for the engineering story.
Global HL7 AI Challenge — Transformative Impact in Healthcare Award (2025)
Let’s Talk Doc
HL7’s official recipient list names Shailesh Dudala; the official winners page ties the award to Let’s Talk Doc. Individual component ownership remains résumé-supported.Official evidence ↗Best of ViVE — User’s Choice Voting (2025)
Team Re-Admit
Public Devpost project and result.Official evidence ↗HiCounselor Diabetes Risk Prediction Challenge (2024)
Diabetes risk prediction
Recognition and model result are résumé-supported; no payer affiliation is asserted.Contact
Building a workflow where an AI answer has to become a defensible action?
Let’s talk about the architecture between those two points.