diff --git a/emfit/__init__.py b/emfit/__init__.py
new file mode 100755
index 0000000..370f926
--- /dev/null
+++ b/emfit/__init__.py
@@ -0,0 +1,370 @@
+#!/usr/bin/env python3
+import json
+import logging
+from collections import OrderedDict as odict
+from dataclasses import dataclass
+from datetime import date, datetime, time, timedelta
+from pathlib import Path
+from typing import Dict, Iterator, List, NamedTuple, Any, cast
+
+import kython
+import pytz
+from kython import cproperty, group_by_key
+
+from cachew import cachew
+
+
+def get_logger():
+    return logging.getLogger('emfit-provider')
+
+timed = lambda f: kython.timed(f, logger=get_logger())
+
+def hhmm(minutes):
+    return '{:02d}:{:02d}'.format(*divmod(minutes, 60))
+
+PATH = Path("/L/backups/emfit")
+
+EXCLUDED = [
+    '***REMOVED***', # pretty weird, detected sleep and HR (!) during the day when I was at work
+    '***REMOVED***',
+
+    '***REMOVED***', # some weird sleep during the day?
+]
+
+AWAKE = 4
+
+Sid = str
+
+# TODO FIXME use tz provider for that? although emfit is always in london...
+_TZ = pytz.timezone('Europe/London')
+
+def fromts(ts) -> datetime:
+    dt = datetime.fromtimestamp(ts)
+    return _TZ.localize(dt)
+
+
+class Mixin:
+    # TODO ugh. tricking mypy...
+    sleep_minutes_emfit: int
+
+    @property
+    # ok, I guess that's reasonable way of defining sleep date
+    def date(self):
+        return self.end.date() # type: ignore[attr-defined]
+
+    @cproperty
+    def time_in_bed(self):
+        return int((self.sleep_end - self.sleep_start).total_seconds()) // 60  # type: ignore[attr-defined]
+
+    @property
+    def recovery(self):
+        return self.hrv_morning - self.hrv_evening  # type: ignore[attr-defined]
+
+    @property
+    def summary(self):
+        return f"""in bed for {hhmm(self.time_in_bed)}
+emfit time: {hhmm(self.sleep_minutes_emfit)}; covered: {self.sleep_hr_coverage:.0f}
+hrv morning: {self.hrv_morning:.0f}
+hrv evening: {self.hrv_evening:.0f}
+avg hr: {self.measured_hr_avg:.0f}
+recovery: {self.recovery:3.0f}
+{self.hrv_lf}/{self.hrv_hf}"""  # type: ignore[attr-defined]
+
+
+# TODO def use multiple threads for that..
+class EmfitOld(Mixin):
+    def __init__(self, sid: str, jj):
+        self.sid = sid
+        self.jj = jj
+
+    def __hash__(self):
+        return hash(self.sid)
+
+    @property
+    def hrv_morning(self):
+        return self.jj['hrv_rmssd_morning']
+
+    @property
+    def hrv_evening(self):
+        return self.jj['hrv_rmssd_evening']
+
+    """
+    Bed time, not necessarily sleep
+    """
+    @property
+    def start(self):
+        return fromts(self.jj['time_start'])
+
+    """
+    Bed time, not necessarily sleep
+    """
+    @property
+    def end(self):
+        return fromts(self.jj['time_end'])
+
+    @property
+    def epochs(self):
+        return self.jj['sleep_epoch_datapoints']
+
+    @property
+    def epoch_series(self):
+        tss = []
+        eps = []
+        for [ts, e] in self.epochs:
+            tss.append(ts)
+            eps.append(e)
+        return tss, eps
+
+    @cproperty
+    def sleep_start(self) -> datetime:
+        for [ts, e] in self.epochs:
+            if e == AWAKE:
+                continue
+            return fromts(ts)
+        raise RuntimeError
+
+    @cproperty
+    def sleep_end(self) -> datetime:
+        for [ts, e] in reversed(self.epochs):
+            if e == AWAKE:
+                continue
+            return fromts(ts)
+        raise RuntimeError
+# 'sleep_epoch_datapoints'
+# [[timestamp, number]]
+
+    # so it's actual sleep, without awake
+    # ok, so I need time_asleep
+    @property
+    def sleep_minutes_emfit(self):
+        return self.jj['sleep_duration'] // 60
+
+    @property
+    def hrv_lf(self):
+        return self.jj['hrv_lf']
+
+    @property
+    def hrv_hf(self):
+        return self.jj['hrv_hf']
+
+    @property
+    def strip_awakes(self):
+        ff = None
+        ll = None
+        for i, [ts, e] in enumerate(self.epochs):
+            if e != AWAKE:
+                ff = i
+                break
+        for i in range(len(self.epochs) - 1, -1, -1):
+            [ts, e] = self.epochs[i]
+            if e != AWAKE:
+                ll = i
+                break
+        return self.epochs[ff: ll]
+
+
+    # # TODO epochs with implicit sleeps? not sure... e.g. night wakeups.
+    # # I guess I could input intervals/correct data/exclude days manually?
+    # @property
+    # def pulse_percentage(self):
+
+    #     # TODO pulse intervals are 4 seconds?
+    #     # TODO ok, how to compute that?...
+    #     # TODO cut ff start and end?
+    #     # TODO remove awakes from both sides
+    #     sp = self.strip_awakes
+    #     present = {ep[0] for ep in sp}
+    #     start = min(present)
+    #     end = max(present)
+    #     # TODO get start and end in one go?
+
+    #     for p in self.iter_points():
+    #         p.ts 
+
+
+    #     INT = 30
+
+    #     missing = 0
+    #     total = 0
+    #     for tt in range(start, end + INT, INT):
+    #         total += 1
+    #         if tt not in present:
+    #             missing += 1
+    #     # TODO get hr instead!
+    #     import ipdb; ipdb.set_trace() 
+    #     return missing
+
+
+    #     st = st[0][0]
+    #     INT = 30
+    #     for [ts, e] in sp:
+    #         if e == AWAKE:
+    #             continue
+    #         return fromts(ts)
+    #     raise RuntimeError
+    #     pass
+
+    def __str__(self) -> str:
+        return f"from {self.sleep_start} to {self.sleep_end}"
+
+# measured_datapoints
+# [[timestamp, pulse, breath?, ??? hrv?]] # every 4 seconds?
+
+    def iter_points(self):
+        for ll in self.jj['measured_datapoints']:
+            [ts, pulse, br, activity] = ll
+            # TODO what the fuck is whaat?? It can't be HRV, it's about 500 ms on average
+            # act in csv.. so it must be activity? wonder how is it measured.
+            # but I guess makes sense. yeah,  "measured_activity_avg": 595, about that
+            # makes even more sense given tossturn datapoints only have timestamp
+            yield ts, pulse
+
+    @property
+    def sleep_hr(self):
+        tss = []
+        res = []
+        for ts, pulse in self.iter_points():
+            if self.sleep_start < fromts(ts) < self.sleep_end:
+                tss.append(ts)
+                res.append(pulse)
+        return tss, res
+
+    @property
+    def sleep_hr_series(self):
+        return self.sleep_hr
+
+    @property
+    def hrv(self):
+        tss = []
+        res = []
+        for ll in self.jj['hrv_rmssd_datapoints']:
+            [ts, rmssd, _, _, almost_always_zero, _] = ll
+            # timestamp,rmssd,tp,lfn,hfn,r_hrv
+            # TP is total_power??
+            # erm. looks like there is a discrepancy between csv and json data.
+            # right, so web is using api v 1. what if i use v1??
+            # definitely a discrepancy between v1 and v4. have no idea how to resolve it :(
+            # also if one of them is indeed tp value, it must have been rounded.
+            # TODO what is the meaning of the rest???
+            # they don't look like HR data.
+            tss.append(ts)
+            res.append(rmssd)
+        return tss, res
+
+    @property
+    def measured_hr_avg(self):
+        return self.jj["measured_hr_avg"]
+
+    @cproperty
+    def sleep_hr_coverage(self):
+        tss, hrs = self.sleep_hr
+        covered = len([h for h in hrs if h is not None])
+        expected = len(hrs)
+        return covered / expected * 100
+
+
+# right, so dataclass is better because you can use mixins
+@dataclass(eq=True, frozen=True)
+class Emfit(Mixin):
+    sid: Sid
+    hrv_morning: float
+    hrv_evening: float
+    start: datetime
+    end  : datetime
+    sleep_start: datetime
+    sleep_end  : datetime
+    sleep_hr_coverage: float
+    measured_hr_avg: float
+    sleep_minutes_emfit: int
+    hrv_lf: float
+    hrv_hf: float
+
+    @classmethod
+    def make(cls, em) -> Iterator['Emfit']:
+        # TODO FIXME res?
+        logger = get_logger()
+        if em.epochs is None:
+            logger.error('%s (on %s) got None in epochs! ignoring', em.sid, em.date)
+            return
+
+        yield cls(**{
+            # pylint: disable=no-member
+            k: getattr(em, k) for k in Emfit.__annotations__
+        })
+
+
+# TODO move to common?
+def dir_hash(path: Path):
+    mtimes = tuple(p.stat().st_mtime for p in sorted(path.glob('*.json')))
+    return mtimes
+
+
+@cachew(cache_path=Path('/L/data/.cache/emfit.cache'), hashf=dir_hash)
+def iter_datas_cached(path: Path) -> Iterator[Emfit]:
+    for f in sorted(path.glob('*.json')):
+        sid = f.stem
+        if sid in EXCLUDED:
+            continue
+
+        em = EmfitOld(sid=sid, jj=json.loads(f.read_text()))
+        yield from Emfit.make(em)
+
+
+def iter_datas(path=PATH) -> Iterator[Emfit]:
+    yield from iter_datas_cached(path)
+
+
+def get_datas() -> List[Emfit]:
+    return list(sorted(iter_datas(), key=lambda e: e.start))
+# TODO move away old entries if there is a diff??
+
+
+@timed
+def by_night() -> Dict[date, Emfit]:
+    logger = get_logger()
+    res: Dict[date, Emfit] = odict()
+    # TODO shit. I need some sort of interrupted sleep detection?
+    for dd, sleeps in group_by_key(get_datas(), key=lambda s: s.date).items():
+        if len(sleeps) > 1:
+            logger.warning("multiple sleeps per night, not handled yet: %s", sleeps)
+            continue
+        [s] = sleeps
+        res[s.date] = s
+    return res
+
+
+
+def test():
+    datas = get_datas()
+    for d in datas:
+        pass
+
+
+def test_tz():
+    datas = get_datas()
+    for d in datas:
+        assert d.start.tzinfo is not None
+        assert d.end.tzinfo is not None
+        assert d.sleep_start.tzinfo is not None
+        assert d.sleep_end.tzinfo is not None
+
+    # https://qs.emfit.com/#/device/presence/***REMOVED***
+    # this was winter time, so GMT, UTC+0
+    sid_20190109 = '***REMOVED***'
+    [s0109] = [s for s in datas if s.sid == sid_20190109]
+    assert s0109.end.time() == time(hour=6, minute=42)
+
+    # summer time, so UTC+1
+    sid_20190411 = '***REMOVED***'
+    [s0411] = [s for s in datas if s.sid == sid_20190411]
+    assert s0411.end.time() == time(hour=9, minute=30)
+
+
+def main():
+    from kython.klogging import setup_logzero
+    setup_logzero(get_logger(), level=logging.DEBUG)
+    for k, v in by_night().items():
+        print(k, v.start, v.end)
+
+if __name__ == '__main__':
+    main()
diff --git a/plot.py b/plot.py
new file mode 100755
index 0000000..87ec6e8
--- /dev/null
+++ b/plot.py
@@ -0,0 +1,26 @@
+# def stats():
+#     for jj in iter_datas():
+#         # TODO fimezone??
+#         # TODOgetinterval on 16 aug -- err it's pretty stupid. I shouldn't count bed exit interval...
+#         st = fromts(jj['time_start'])
+#         en = fromts(jj['time_end'])
+#         tfmt = "%Y-%m-%d %a"
+#         tot_mins = 0
+#         res = []
+#         res.append(f"{st.strftime(tfmt)} -- {en.strftime(tfmt)}")
+#         for cls in ['rem', 'light', 'deep']:
+#             mins = jj[f'sleep_class_{cls}_duration'] // 60
+#             res += [cls, hhmm(mins)]
+#             tot_mins += mins
+#         res += ["total", hhmm(tot_mins)]
+#         print(*res)
+
+
+# then I could enable emfit feed and slog feed (pulled from all org notes) and see the correlation? also could pull workouts provider (and wlog) -- actually wlog processing could be moved to timeline too
+
+# TODO maybe rmssd should only be computed if we have a reasonable chunk of datas
+
+# TODO X axes: show hours and only two dates
+# TODO 4 is awake, 3 REM, 2 light, 1 deep
+
+# deviartion beyond 25-75 or 75-25 is bad??