Case studies

Case Studies

AI in the real world.

20+

Wards monitored

Hours → minutes

Time to insight

6 in 5 weeks

Releases

Government · Political intelligence

Challenge: Field sentiment across dozens of wards was scattered across social platforms and WhatsApp groups. Leadership had no evidence-based view of where support was shifting.
Approach: We built Sanjaya: a crawl-score-explain loop that treats every ward as a monitored signal, with negative sentiment surfaced first and plain-language Q&A over the evidence.
Implementation: Batch crawling with resumable sweeps, LLM scoring with analyst correction, a Ward Exchange replay view, voice-enabled Ask Sanjaya and draft-and-approve voter replies — delivered over six iterative releases.
Outcome: Daily briefings moved from anecdote to evidence. Teams could see ward trajectories, drill into causes and respond within hours.

Next.jsPrismaAnthropicGeminiElevenLabs

< 3 min

Time to file

None

Literacy required

~100%

Data completeness

Government · Citizen services

Challenge: Citizens queued for hours to file complaints on paper forms in a language they did not read. Data quality was poor and follow-up impossible.
Approach: We replaced the form with a conversation. The citizen speaks once; the system detects the language, extracts the structured complaint, confirms in their language and issues a tracked receipt.
Implementation: iPad kiosk-mode PWA, provider admin for LLM/ASR/TTS, structured extraction pipeline, Hindi-safe PDF generation and a hashtag sentiment utility for the same office.
Outcome: Filing time fell below three minutes with no literacy requirement. Every grievance now carries clean structured data.

Next.jsSarvamGoogle TTSOpenRouter

~3s

Latency

Providers

< 1 hour

Operator training

Media · Live reporting

Challenge: Live sessions are long, multilingual and hard to follow. Audiences and media teams lacked a readable real-time companion.
Approach: The Live Reporting Desk ingests the stream, transcribes with automatic language detection, overlays the active speaker and gives operators full control of providers and translation.
Implementation: Streaming ASR with provider choice, operator-managed speaker cards, speaker directory with CSV import, and LAN-accessible deployment for in-house use.
Outcome: A live, searchable transcript companion with ~3 second latency, operable by non-technical staff.

Next.jsFastAPISarvamGemini

Start here

Have an AI problem worth solving?

Let's turn it into something real.