Work / AI · Agents / RealtyAssistant
No. 25 · Live · Real Estate Lead QualificationReal estate sales lose 70% of leads to follow-up failure. The leads call at 11 PM in code-mixed Indian English. Sales reps speak office English by day only. RealtyAssistant picks up on the first ring, handles Noida-Gurugram accents, and never drops a lead, even when the database breaks.
Act I · The Lost Leads
The phone rings after dinner, after the kids, after the Zoom calls. The agent's office-hour answering service takes a name. By morning the lead has already toured a competitor's project.
"Bhai 3BHK chahiye Noida Extension mein, around 50 lakh, schools nearby, RERA approved, kal site visit ho sakta hai?" Most voice bots stop existing at sentence two.
The lead asks four questions. The CRM gets two. Sales calls back twice. The lead has moved on. Industry-standard outcome: lost.
Most "AI receptionist" platforms price for US enterprises. The mid-market broker in Gurugram cannot afford the wrapper, let alone the AI.
Act II · The Promise
Lead
Hi, I am looking for a 3 BHK apartment in Noida Extension, around 50 lakhs.
3 BHKApartmentNoida Extension≤ ₹50L
RealtyAssistant
Got it. Three-BHK in Noida Extension under 50 lakhs. One quick check, RERA-approved projects only, or are open-plot too?
Lead
RERA approved, school nearby, possession within 1 year.
RERASchool proximityPossession ≤ 12mo
RealtyAssistant
Pulling live listings from realtyassistant.in. Six matches. I will text you the top three with site-visit slots for tomorrow. What time works?
Lead
11 baje free hoon, brother.
11:00 ISTSite visit confirmed
Act III · The Fail-Safe
A ₹500/month VPS will fail. The disk will fill. The LLM will time out. The hybrid architecture treats every dependency as untrusted. Leads are captured, no matter what.
Layer 01 · Voice
Local Whisper engine for transcription. Fuzzy-matching for Indian place names: Noida, Greater Noida, Noida Extension, Sector 150 are all canonicalised to one entity.
Scenario: lead says "Noyda" → matched → entity captured
Layer 02 · Brain
Local Ollama serves the common case. If the local model times out or hallucinates, Gemini takes over. If both fail, deterministic regex extraction still captures the structured fields.
Scenario: Ollama hangs → Gemini in 1.2s → lead saved
Layer 03 · Memory
Auto-detects database corruption, runs integrity check, repairs in place. If the disk is unwritable, falls back to an in-memory store and writes to log so leads can be replayed.
Scenario: disk full → in-memory + log → zero loss
Layer 04 · Surface
Scrapes realtyassistant.in in real time so the listings shown to the lead are the listings still on market. Cached for 60 seconds, refreshed in the background.
Scenario: scraper down → cached pool served
Act IV · One-Click Deploy
Checks Python 3.10+, creates venv, installs deps, frees port 20000, starts the server. Open localhost:20000/demo and start qualifying.
Installs system deps, sets up nginx reverse-proxy, registers as a systemd service, comes back up after reboot. Total cost: one ₹500/month VPS.
Act V · Proof
Single Ubuntu VPS. Nginx reverse-proxy. Local Ollama for the common case. Gemini for the hard cases. Cheaper than a single junior outbound call.
Place names like Noyda, NOIDA, Noida-Ext, Greater Noida West all normalise to canonical entities. Pricing tokens like "50 lakh" and "5 crore" parsed at extraction time.
Pluggable voice provider via VOICE_PROVIDER env. Switch from Twilio to VAPI without code changes. Webhook-driven, scales horizontally.
Every captured lead, every extracted entity, every transcript chunk. SQLite-backed, exportable. Sales sees what the AI saw.
The Stack
I build voice-first products for India that survive bad disks, bad networks and code-mixed sentences. If your CX is leaking leads, talk to me.