Tathastha Labs
Engineering case studies

Six systems, one pattern: turning messy human input into structured output.

Ambient clinical documentation, agentic planning, a natural-language data agent, a grounded research assistant, computer-vision measurement, and two fast feasibility demos — each one built around the same core move: take unstructured input (speech, photos, documents, activity logs) and turn it into something structured, searchable, or actionable.

6 case studies 4 domains — clinical care, planning, research, computer vision 12+ distinct AI techniques
Production
Healthcare · Ambient clinical intelligence

AI Clinical Documentation & EHR Copilot Suite

Clinicians talk to their patients like normal. The note writes itself, in whichever EHR the clinic already runs.

Read case study arrow_forward
Production
Productivity · Personal & team planning

Agentic Goal-Planning & OKR Execution Platform

Describe what you want, answer a few questions, and walk away with a real, trackable plan of goals and tasks.

Read case study arrow_forward
Evolving
Operations · Automated reporting

Automated Progress-Intelligence & MCP Data-Agent Pipeline

A plain-English "here's where things stand" lands every morning — no one has to write a report or run a query.

Read case study arrow_forward
Prototype
Health & wellness · Grounded scientific Q&A

Multi-Agent Biomarker Research Assistant

Ask what a lab result means and get an answer grounded in real research — not a generic guess.

Read case study arrow_forward
Proof of concept
Health & wellness · Remote monitoring via computer vision

Computer-Vision Wound Assessment Tool

Take a photo. Get back a measured size, an estimated depth, and a tissue-condition breakdown.

Read case study arrow_forward
Proof of concept
Healthcare · Feasibility demos

Rapid Prototyping: Clinical Notes from Voice and Text

Two fast, cheap demos, built to prove an idea works before committing real engineering time to it.

Read case study arrow_forward
Production Healthcare · Ambient clinical intelligence

AI Clinical Documentation & EHR Copilot Suite

Clinicians talk to their patients like normal. The note writes itself, in the system they already use.

micLive consult graphic_eqStreaming ASR smart_toyLLM note + EHR sync fact_checkSigned chart note

The problem

Doctors spend more time typing than treating. Every extra minute charting is a minute not spent with a patient — and most clinics can't or won't change the record system they've used for years.

What we built

The suite centers on a real-time visit copilot that listens during a consult and drafts a structured note as the conversation happens, plus a browser extension that injects that copilot directly into the clinician's existing EHR tab rather than asking them to open a separate app. Behind it, an EHR-connect layer normalizes OAuth and data mapping across a dozen different EHR platforms, and a serverless, event-driven pipeline handles asynchronous note generation whenever a recording is uploaded outside the live-visit flow. Patient-facing intake and messaging apps extend the same backend to the pre- and post-visit parts of the workflow.

Architecture & AI techniques

Real-time streaming ASR LLM structured note generation LLM-based speaker diarization Speech-to-speech interpretation ~10-EHR OAuth integration layer Pluggable multi-vendor LLM/ASR layers

Value delivered

The direct effect is fewer minutes per visit spent typing and more spent with the patient — the note is drafted automatically, in the clinician's own EHR, in the note format their specialty already uses.

Because the integration layer is vendor-agnostic, a clinic doesn't need to change record systems, and the product isn't locked to one AI vendor if pricing, quality, or availability shifts. Time-saved and adoption figures are intentionally not stated here — they should only be published from real, independently verified usage data.

Production Productivity · Personal & team planning

Agentic Goal-Planning & OKR Execution Platform

Describe what you want. Answer a few questions. Walk away with a real plan — goals, tasks, and a way to track them weekly.

flagStated goal forumClarifying Q&A agent account_treeObjective + task scoring checklistTrackable weekly plan

The problem

Most people know the outcome they want — get fit, launch a project, hit a revenue number — but stall at turning it into concrete steps. Teams have the same problem at a bigger scale: strategy exists on a slide, but nobody's clear on this week's tasks.

What we built

A conversational agent asks a small, fixed set of clarifying questions about the user's goal, proposes 3–5 high-level objectives for confirmation, then expands confirmed objectives into a full tree of goals, sub-goals, and tasks — each scored on effort, impact, and control. That plan is tracked weekly with progress and risk indicators, streaks, and a daily planning view, inside a multi-tenant product that also handles teams, coaching roles, and subscription billing.

Architecture & AI techniques

LangGraph multi-agent supervisor Conversational plan generation Effort/impact/control scoring Multi-tenant SaaS (schema-per-tenant) Vector-based template retrieval (in progress)

Value delivered

What used to take a working session with a coach or a blank page and good intentions now takes one guided conversation. The output isn't generic advice — it's a structured, trackable plan with tasks the user (or team) can act on the same day.

For teams, the same engine scales from one person's personal goal to an organization's strategic goals, with the accountability layer — weekly tracking, coaching, risk flags — built in rather than bolted on.

Evolving Operations · Automated reporting & conversational data access

Automated Progress-Intelligence & MCP Data-Agent Pipeline

Every morning, a plain-English "here's where things stand" lands for every person, goal, and team — with no one having to write a report or run a query.

datasetDaily activity data hubRate-limited LLM fan-out chatMCP data agent summarizePlain-English digest

The problem

With activity spread across every goal, task, and team, leaders either read everything themselves or read nothing. Neither scales, and by the time a weekly report gets written, it's already stale.

What we built

A scheduled orchestration job runs every 30 minutes, fans out across every user, goal, organization, and team, and asks an LLM to summarize recent activity and flag risks — with batching, rate limiting, retries, and token accounting so it doesn't fall over at scale. A second version adds a delivery step, formatting the day's summaries into a report. Separately, a Model Context Protocol (MCP) server exposes the underlying database as a safe, read-only query tool — restricted to an explicit table allowlist and SELECT-only access — paired with an agent client that adds lightweight retrieval over past summaries so it can answer follow-up questions with memory of prior context.

Architecture & AI techniques

Scheduled multi-entity summarization Model Context Protocol (MCP) Rate-limited, idempotent fan-out Tool-calling agent with memory RAG-lite over historical summaries

Value delivered

Instead of a leader hunting through dashboards or waiting for a weekly report, a concise, plain-English summary is simply waiting for them each day — for themselves, for a team, or for the organization as a whole.

The natural-language data agent points toward a future where anyone can ask a question about progress directly, instead of every stakeholder needing their own custom report built for them. The read-only data-agent layer is currently a proof of concept, not yet a hardened service.

Prototype Health & wellness · Grounded scientific Q&A

Multi-Agent Biomarker Research Assistant

Ask what a lab result means and get an answer grounded in real medical literature — not a generic guess, and not a search you have to interpret yourself.

helpLab-result question alt_routeIntent routing manage_searchHybrid RAG + self-critique verifiedCited grounded answer

The problem

A biomarker result like HbA1c or TSH comes back with a number and a reference range, and most people are left to search the internet themselves — with no way to know if what they're reading is accurate or relevant to them.

What we built

A supervisor agent first classifies each incoming question — small talk, a question about the user's own profile, or a scientific question about a biomarker — and routes accordingly. Profile questions go to a small, isolated agent that answers strictly from the user's own stored data. Scientific questions go to a retrieval agent that searches a structure-aware index of biomedical research papers, drafts an answer, scores that draft against a fitness function, and — if the score falls short — refines and re-synthesizes before responding, rather than returning its first-pass answer.

Architecture & AI techniques

Intent classification / agent routing Structure-aware chunking Hybrid keyword + vector retrieval Self-critique generate→score→refine loop ReAct tool-calling retrieval agent

Value delivered

Instead of an AI assistant confidently making something up about a health topic, this one is built to know the difference between "I know this about you" and "here's what the research says" — and to double-check its own scientific answers before giving them. This is a pre-production system: it has not been clinically validated, and no accuracy or outcome claims should be attached to it without that validation first.

Proof of concept Health & wellness · Remote monitoring via computer vision

Computer-Vision Wound Assessment Tool

Take a photo. Get back a measured size, an estimated depth, and a breakdown of tissue condition — the kind of read a clinician would normally need to be in the room to give.

photo_cameraWound photo content_cutInstance segmentation square_footPCA measurement + depth monitoringStructured wound metrics

The problem

Monitoring a healing wound — a pressure sore, a diabetic ulcer — usually means an in-person visit just to measure it. For homebound or remote patients, that's a real access barrier.

What we built

An uploaded photo is run through a custom-trained segmentation model to isolate the wound region from surrounding skin. Principal-axis analysis on that mask produces length, width, and area, calibrated against a reference object visible in frame. The cropped wound region is separately passed through a monocular depth-estimation model to produce a relative depth estimate, and a color-space analysis buckets the visible tissue into categories associated with different stages of healing.

Architecture & AI techniques

Custom-trained instance segmentation PCA-based geometric measurement Monocular depth estimation Rule-based color classification

Value delivered

The core idea — structured wound measurements from a phone photo, without requiring an in-person visit — is demonstrated and working end-to-end. Important caveat: depth is a relative, not an absolute, measurement, and size calibration currently assumes a fixed reference object rather than detecting one in the photo. This should be treated as a promising prototype, not a clinically validated measurement tool, until those two points are addressed and independently verified.

Proof of concept Healthcare · Feasibility demos

Rapid Prototyping: Clinical Notes from Voice and Text

Two quick demonstrations, built to prove the idea works before investing in a full product: one turns existing session notes into a formatted clinical note, the other turns a spoken conversation into one.

descriptionText or spoken visit graphic_eqASR + diarization smart_toyTemplate-matched prompting fact_checkFormatted note draft

The problem

Before committing real engineering time, it's worth proving an idea works on a small scale — these two demos exist to answer "can this actually work" quickly and cheaply.

What we built

The first demo takes an uploaded document, extracts its text, and prompts an LLM with a note-format-specific template — SOAP, DAP, BIRP, and several others — to produce a formatted clinical note. The second takes an uploaded audio recording, transcribes it with an automatic speech recognition model, optionally labels speaker turns, and then runs the same kind of prompt-templated note generation over the resulting transcript.

Architecture & AI techniques

Prompt-engineered note generation Automatic speech recognition Speaker diarization Multi-format template library (SOAP/DAP/BIRP)

Value delivered

Both ideas were validated quickly and cheaply, which is exactly the point of a demo: prove the concept works before spending real engineering budget on a hardened pipeline. Both fed directly into the decision to invest in the fuller async note-generation pipeline described in our clinical documentation suite above.

Have a similar problem?

Let's talk about what you're trying to build.

Send the one paragraph you'd normally write to a colleague. We'll reply within two working days with whether we think we can help — and what we'd do first if we could.

Start a conversation

Product and internal system names referenced above have been generalized. Maturity labels reflect an honest, code-level read of each system, and metrics such as adoption, time saved, or accuracy are not claimed unless independently validated.